Sie sind auf Seite 1von 41

ULTRAMETRIC SUBSETS WITH LARGE HAUSDORFF DIMENSION

MANOR MENDEL AND ASSAF NAOR

Abstract. It is shown that for every (0, 1), every compact metric space (X, d) has a
compact subset S X that embeds into an ultrametric space with distortion O(1/), and
dimH (S) > (1 ) dimH (X),
where dimH () denotes Hausdorff dimension. The above O(1/) distortion estimate is shown
to be sharp via a construction based on sequences of expander graphs.

Contents
1. Introduction 2
1.1. An overview of the proof of Theorem 1.5 4
1.2. The low distortion regime 5
1.3. Further applications 6
1.3.1. Urbanskis problem 6
1.3.2. Talagrands majorizing measures theorem 7
2. Reduction to finite metric spaces 8
3. Combinatorial trees and fragmentation maps 11
4. From fragmentation maps to covering theorems 13
5. Asymptotically optimal fragmentation maps: proof of Lemma 4.3 15
6. An intermediate fragmentation map: proof of Lemma 5.5 22
7. The initial fragmentation map: proof of Lemma 6.2 25
8. An iterated Holder argument for trees: proof of Lemma 6.5 27
9. Proof of Theorem 5.3 33
10. Impossibility results 35
10.1. Trees of metric spaces 35
10.2. Expander fractals 39
10.3. G (n, 1/2) fractals 39
References 40

2010 Mathematics Subject Classification. 30L05,46B85,37F35.


Key words and phrases. Bi-Lipschitz embeddings, Hausdorff dimension, ultrametrics, Dvoretzkys
theorem.
M. M. was partially supported by ISF grants 221/07 and 93/11, BSF grants 2006009 and 2010021, and
a gift from Cisco Research Center. A. N. was partially supported by NSF grant CCF-0832795, BSF grants
2006009 and 2010021, and the Packard Foundation. Part of this work was completed when M. M. was
visiting Microsoft Research and University of Washington, and A. N. was visiting the Discrete Analysis
program at the Isaac Newton Institute for Mathematical Sciences and the Quantitative Geometry program
at the Mathematical Sciences Research Institute.

1
1. Introduction
Given D > 1, a metric space (X, dX ) is said to embed with distortion D into a metric
space (Y, dY ) if there exists f : X Y and > 0 such that for all x, y X we have
dX (x, y) 6 dY (f (x), f (y)) 6 DdX (x, y). (1)
Note that when Y is a Banach space the scaling factor can be dropped in the definition (1).
Answering positively a conjecture of Grothendieck [15], Dvoretzky proved [13] that for
every k N and D > 1 there exists n = n(k, D) N such that every n-dimensional normed
space has a k-dimensional linear subspace that embeds into Hilbert space with distortion D;
see [28, 27, 32] for the best known bounds on n(k, D).
Bourgain, Figiel and Milman [9] studied the following problem as a natural nonlinear
variant of Dvoretzkys theorem: given n N and D > 1, what is the largest m N such
that any finite metric space (X, d) of cardinality n has a subset S X with |S| > m
such that the metric space (S, d) embeds with distortion D into Hilbert space? Denote this
value of m by R(n, D). Bourgain-Figiel-Milman proved [9] that for all D > 1 there exists
c(D) (0, ) such that R(n, D) > c(D) log n, and that R(n, 1.023) = O(log n). Following
several investigations [18, 8, 3] that were motivated by algorithmic applications, a more
complete description of the Bourgain-Figiel-Milman phenomenon was obtained in [5].
Theorem 1.1 ([5]). For D (1, ) there exist c(D), c0 (D) (0, ) and (D), 0 (D) (0, 1)
such that for every n N,
if D (1, 2) then c(D) log n 6 R(n, D) 6 c0 (D) log n,
0
if D (2, ) then n1(D) 6 R(n, D) 6 n1 (D) .
Highlighting the case of large D, which is most relevant for applications, we have the
following theorem.
Theorem 1.2 ([5, 25, 30]). For every (0, 1) and n N, any n-point metric space (X, d)
has a subset S X with |S| > n1 that embeds into an ultrametric space with distortion
2e/. On the other hand, there exists a universal constant c > 0 with the following property.
For every n N there is an n-point metric space Xn such that for every (0, 1) all subsets
Y X with |Y | > n1 incur distortion at least c/ in any embedding into Hilbert space.
Recall that a metric space (U, ) is called an ultrametric space if for every x, y, z U we
have (x, y) 6 max {(x, z), (z, y)}. Any separable ultrametric space admits an isometric
embedding into Hilbert space [39]. Hence the subset S from Theorem 1.2 also embeds with
the stated distortion into Hilbert space, and therefore Theorem 1.2 fits into the Bourgain-
Figiel-Milman framework. Note, however, that the stronger statement that S embeds into
an ultrametric space is needed for the applications in [5, 25], and that the matching lower
bound in Theorem 1.2 is for the weaker requirement of embeddability into Hilbert space.
Thus, a byproduct of Theorem 1.2 is the assertion that, in general, the best way (up to
constant factors) to find a large approximately Euclidean subset is to actually find a subset
satisfying the more stringent requirement of being almost ultrametric. The existence of
the metric spaces {Xn } n=1 from Theorem 1.2 was established in [5]. The estimate 2e/ on
the ultrametric distortion of the subset S from Theorem 1.2 is due to [30], improving by a
constant factor over the bound from [25], which itself improves (in an asymptotically optimal
way) on the distortion bound of O (1 log(2/)) from [5].

2
In what follows, dimH (X) denotes the Hausdorff dimension of a metric space X. Inspired
by the above theorems, Terence Tao proposed (unpublished, 2006) another natural variant
of the nonlinear Dvoretzky problem: one can keep the statement of Dvoretzkys theorem
unchanged in the context of general metric spaces, while interpreting the notion of dimension
in the appropriate category. Thus one arrives at the following question.
Question 1.3 (The nonlinear Dvoretzky problem for Hausdorff dimension). Given > 0
and D > 1, what is the supremum over those > 0 with the following property. Every
compact metric space X with dimH (X) > has a subset S X with dimH (S) > that
embeds into Hilbert space with distortion D?
The restriction of Question 1.3 to compact metric spaces is not severe. For example, if
X is complete and separable then one can first pass to a compact subset of X with the
same Hausdorff dimension, and even the completeness of X can be replaced by weaker
assumptions; see [11, 16]. We will not address this issue here and restrict our discussion to
compact metric spaces, where the crucial subtleties of the problem are already present.
Our purpose here is to provide answers to Question 1.3 in various distortion regimes, the
main result being the following theorem.
Theorem 1.4. There exists a universal constant C (0, ) such that for every (0, 1)
and (0, ), every compact metric space X with dimH (X) > has a closed subset S X
with dimH (S) > (1 ) that embeds with distortion C/ into an ultrametric space. In the
reverse direction, there is a universal constant c > 0 such that for every > 0 there exists a
compact metric space X with dimH (X ) = such that if S X satisfies dimH (S) > (1)
then S incurs distortion at least c/ in any embedding into Hilbert space.
The construction of the spaces X from Theorem 1.4 builds on the examples of [5], which
are based on expander graphs. The limiting spaces X obtained this way can therefore be
called expander fractals; their construction is discussed in Section 10.
Our main new contribution leading to Theorem 1.4 is the following structural result for
general metric measure spaces. In what follows, by a metric measure space (X, d, ) we mean
a compact metric space (X, d), equipped with a Borel measure such that (X) < . For
r > 0 and x X, the corresponding closed ball is denoted B(x, r) = {y X : d(x, y) 6 r}.
Theorem 1.5. For every (0, 1) there exists c (0, ) with the following property.
Every metric measure space (X, d, ) has a closed subset S X such that (S, d) embeds into
an ultrametric space with distortion 9/, and for every {xi }iI X and {ri }iI [0, )
such that the balls {B(xi , ri )}iI cover S, i.e.,
[
B(xi , ri ) S, (2)
iI

we have X
(B(xi , c ri ))1 > (X)1 . (3)
iI

Theorem 1.5 contains Theorem 1.2 as a simple special case. Indeed, consider the case
when X is finite, say |X| = n, the measure is the counting measure, i.e., (A) = |A|
for all A X, and all the radii {ri }iI vanish. In this case B(xi , ri ) = B(xi , c ri ) = {xi },
and therefore the covering condition (2) implies that {xi }iI S. Inequality (3) therefore

3
implies that |S| > n1 , which is the (asymptotically sharp) conclusion of Theorem 1.2, up
to a constant multiplicative factor in the distortion.
Theorem 1.5 also implies Theorem 1.4. To see this assume that (X, d) is a compact
metric space and dimH (X) > . The Frostman lemma (see [16] and [24, Ch. 8]) implies
that there exists a constant K (0, ) and a Borel measure such that (X) > 0 and
(B(x, r)) 6 Kr for all r > 0 and x X. An application of Theorem 1.5 to the metric
measure space (X, d, ) yields a closed subset S X that embeds into an ultrametric space
with distortion O(1/) and satisfies the covering condition (3). Thus, all the covers of S by
balls {B(xi , ri )}iI satisfy
X X
(X)1 6 (B(xi , c ri ))1 6 (Kc ri )1 .
iI iI

Hence,
X (1) (X)1
ri > (1)
.
iI K 1 c
This means that the (1 )-Hausdorff content1 of S satisfies
(1) (X)1
H (S) > (1)
> 0,
K 1 c


and therefore dimH (S) = inf > 0 : H (S) = 0 > (1 ), as asserted in Theorem 1.4.
To summarize the above discussion, the general structural result for metric measure spaces
that is contained in Theorem 1.5 implies the sharp Bourgain-Figiel-Milman style nonlinear
Dvoretzky theorem when applied to trivial covers of S by singletons. The nonlinear Dvoret-
zky problem for Hausdorff dimension is more subtle since one has to argue about all possible
covers of S, and this is achieved by applying Theorem 1.5 to the metric measure space in-
duced by a Frostman measure. In both of these applications the value of the constant c in
Theorem 1.5 is irrelevant, but we anticipate that it will play a role in future applications of
2
Theorem 1.5. Our argument yields the bound c = eO(1/ ) , but we have no reason to believe
that this dependence on is optimal. We therefore pose the following natural problem.
Question 1.6. What is the asymptotic behavior as 0 of the best possible constant c
in Theorem 1.5?
1.1. An overview of the proof of Theorem 1.5. Theorem 1.1 was proved in [5] via a
deterministic iterative construction of a sufficiently large almost ultrametric subset S of a
given finite metric space (X, d). In contrast, Theorem 1.2 was proved in [25] via a significantly
shorter probabilistic argument. It is shown in [25] how to specify a distribution (depending
on the geometry of X) over random subsets S X that embed into an ultrametric space with
small distortion, yet their expected cardinality is large. The lower bound on the expected
cardinality of S is obtained via a lower bound on the probability Pr [x S] for each x X.
Such a probabilistic estimate seems to be quite special, and we do not see how to argue
probabilistically about all possible covers of a random subset S, as required in Theorem 1.5.
In other words, a reason why Question 1.3 is more subtle than the Bourgain-Figiel-Milman
problem is that ensuring that S is large is in essence a local requirement, while ensuring that
1Recall
that for > 0 the -Hausdorff content of a metric space (Z, d) is defined to be the infimum of
P
r
jJ j over all possible covers of Z by balls {B(zj , rj )}jJ ; see [24].

4
S is high-dimensional is a global requirement: once S has been determined one has to argue
about all possible covers of S rather than estimating Pr [x S] for each x X separately.
For the above reason our proof of Theorem 1.5 is a deterministic construction which uses in
some of its steps adaptations of the methods of [5], in addition to a variety of new ingredients
that are needed in order to handle a covering condition such as (3). Actually, in order to
obtain the sharp O(1/) distortion bound of Theorem 1.5 we also use results of [25, 30] (see
Theorem 9.1 below), so in fact Theorem 1.5 is based on a combination of deterministic and
probabilistic methods, the deterministic steps being the most substantial new contribution.
The proof of Theorem 1.5 starts with a reduction of the problem to the case of finite metric
spaces; see Section 2. Once this is achieved, the argument is a mixture of combinatorial,
analytic and geometric arguments, the key objects of interest being fragmentation maps.
These are maps that are defined on rooted combinatorial trees and assign to each vertex of
the tree a subset of the metric space (X, d) in a way that respects the tree structure, i.e.,
the set corresponding to an offspring of a vertex is a subset of the set corresponding to the
vertex itself, and vertices lying on distinct root-leaf paths are assigned to disjoint subsets of
X. We also require that leaves are mapped to singletons.
Each fragmentation map corresponds to a subset of X (the images of the leaves), and our
goal is to produce a fragmentation map that corresponds to a subset of X which satisfies the
conclusion of Theorem 1.5. To this end, we initiate the iteration via a bottom-up construction
of a special fragmentation map; see Section 7. We then proceed to iteratively prune (or
sparsify) this initial tree so as to produce a smaller tree whose leaves satisfy the conclusion
of Theorem 1.5. At each step we argue that there must exist sufficiently many good pruning
locations so that by a pigeonhole argument we can make the successive pruning locations
align appropriately; see Section 8.
At this point, the subset corresponding to fragmentation map that we constructed is
2
sufficient to prove Theorem 1.5 with a weaker bound of eO(1/ ) on its ultrametric distortion;
see Remark 5.6. To get the optimal distortion we add another pruning step guided by a
weighted version (proved in Section 9) of the nonlinear Dvoretzky theorem for finite metric
spaces. The mechanism of this second type of pruning is described in Section 5.
It is impossible to describe the exact details of the above steps without introducing a
significant amount of notation and terminology, and specifying rather complicated inductive
hypotheses. We therefore refer to the relevant sections for a detailed description. To help
motivate the lengthy arguments, in the body of this paper we present the proof in a top-down
fashion which is opposite to the order in which it was described above.
1.2. The low distortion regime. We have thus far focused on Question 1.3 in the case of
high distortion embeddings into ultrametric spaces. There is also a Hausdorff dimensional
variant of the phase transition at distortion 2 that was described in Theorem 1.1.
Theorem 1.7 (Distortion 2 + ). There exists a universal constant c (0, ) such that
for every (0, 1/2), any compact metric space (X, d) of finite Hausdorff dimension has a
closed subset S X that embeds with distortion 2 + in an ultrametric space, and
c
dimH (S) > dimH (X). (4)
log(1/)
For distortion strictly less than 2 the following theorem shows that there is no nonlinear
Dvoretzky phenomenon in terms of Hausdorff dimension.

5
Theorem 1.8. For every (0, ) there exists a compact metric space (X, d) of Hausdorff
dimension , such that if S X embeds into Hilbert space with distortion strictly smaller
than 2 then dimH (S) = 0.
It was recently observed in [14] that Theorem 1.8 easily implies the following seemingly
stronger assertion: there exists a compact metric space X such that dimH (X ) = , yet
every subset S X that embeds into Hilbert space with distortion strictly smaller than 2
must have dimH (S) = 0.
As in the case of the finite nonlinear Dvoretzky theorem, Question 1.3 at distortion 2
remains open. In the same vein, the correct asymptotic dependence on in (4) is unknown.
Theorem 1.7 follows from the following result in the spirit of Theorem 1.5, via the same
Frostman measure argument.
Theorem 1.9. There exists a universal constant c (0, ) such that for every (0, 1/2)
there exists c0 (0, ) with the following property. Every metric measure space (X, d, )
has a closed subset S X such that (S, d) embeds into an ultrametric space with distortion
2 + , and for every {xi }iI X and {ri }iI [0, ) such that the balls {B(xi , ri )}iI cover
S, we have X c c
(B(xi , c0 ri )) log(1/) > (X) log(1/) . (5)
iI

1.3. Further applications. Several applications of our results have been recently discov-
ered. We discuss some of them here as an indication of how Theorem 1.5 could be used.
Before doing so, we note the obvious observation that since Theorem 1.5 implies Theorem 1.2
it automatically inherits its applications in theoretical computer science. Algorithmic appli-
cations of nonlinear Dvoretzky theory include the best known lower bound for the randomized
k-server problem [4, 5], and the design of a variety of proximity data structures [25], e.g.,
the only known approximate distance oracles with constant query time, improving over the
important work of Thorup and Zwick [37] (this improvement is sharp [33, 41]. Nonlinear
Dvoretzky theory is the only known method to produce such sharp constructions).
A version of Theorem 1.4 in the case of infinite Hausdorff dimension, in which the con-
clusion is that S is also infinite dimensional, was recently obtained in [14]. In the ensuing
subsections we discuss two additional research directions: surjective cube images of spaces
with large Hausdorff dimension, and the majorizing measures theorem.
1.3.1. Urbanskis problem. This application of Theorem 1.4 is due to Keleti, Mathe and
Zindulka [20]. Urba nski asked [38] whether given n N every metric space (X, d) with
dimH (X) > n admits a surjective Lipschitz map f : X [0, 1]n (Urba nski actually needed
a weaker conclusion). Keleti, Mathe and Zindulka proved that without further assumptions
on X the Urba nski problem has a negative answer, yet if X is an analytic subset of a Polish
space then one can use Theorem 1.4 to solve Urba nskis problem positively. To see how this
can be proved using Theorem 1.4, note that by [16] it suffices to prove this statement when
X is compact. Choose (0, 1) such that dimH (X) > n/(1 ). By Theorem 1.4 there
exists a compact S X with dimH (S) > n, an ultrametric space (U, ), and a bijection
f : S U satisfying d(x, y) 6 (f (x), f (y)) 6 9 d(x, y) for all x, y S.
Since (U, ) is a compact ultrametric space, there exists a linear ordering 6 of U such that
for every a, b U with a 6 b, the order interval [a, b] = {c U : a 6 c 6 b} is a Borel set
satisfying diam([a, b]) = (a, b). This general property of ultrametric spaces follows directly

6
from the well-known representation of such spaces as ends of trees (see [17]); the desired
ordering is then a lexicographical order associated to the tree structure. Since dimH (U ) > n,
we can consider a Frostman probability measure on U , i.e., a Borel measure on U satisfying
(U ) = 1 such that there exists K (0, ) for which (A) 6 K(diam(A))n for all A U .
Define g : U [0, 1] by g(a) = ({x U : x < a}). If a, b U satisfy a < b then
|g(b) g(a)| = ([a, b)) 6 K(diam([a, b]))n = K(a, b)n . Thus g is continuous, implying that
g(U ) = [0, 1] (U is compact and g cannot have any jumps because is atom-free).
Let P : [0, 1] [0, 1]n be a Peano curve (see e.g. [31]), i.e., P ([0, 1]) = [0, 1]n and we have
the 1/n-Holder estimate kP (s) P (t)k2 6 L|s t|1/n for all s, t [0, 1]. Then the mapping
= P g f : S [0, 1]n is surjective and 9K 1/n L/-Lipschitz. There exists : X [0, 1]n
that extends and is CK 1/n L/2 -Lipschitz, where C is a universal constant. This follows
from the absolute extendability property of ultrametric spaces, or more generally metric
trees; see [22]. Alternatively, one can use the nonlinear Hahn-Banach theorem [7, Lem. 1.1],
in which case C will depend on n. Since (X) (S) = [0, 1]n , this concludes the proof of
the Keleti-Mathe-Zindulka positive solution of Urba nskis problem.
The conclusion of Urba nskis problem is known to fail if we only assume that X has
positive n-dimensional Hausdorff measure; see [40], [19] and [2, Thm. 7.4]. However, in the
special case when X is a subset of Rn of positive Lebesgue measure, a well-known conjecture
of Laczkovich [21] asks for the same conclusion, i.e., that there is a surjective Lipschitz
mapping from X onto [0, 1]n . The Laczkovich conjecture has a positive answer [23, 1] when
n = 2, and there is recent exciting (still unpublished) progress on the Laczkovich question for
n > 3 due to Marianna Csornyei and Peter Jones. Note that the above argument implies that
if (X, d) is compact and dimH (X) = n then for every (0, 1) there exists a (1 )-Holder
mapping from X onto [0, 1]n .
1.3.2. Talagrands majorizing measures theorem. Given a metric space (X, d) let PX be the
Borel probability measures on X. The Fernique-Talagrand 2 functional is defined as follows.
Z s  
1
2 (X, d) = inf sup log dr.
PX xX 0 (B(x, r))
In 1987 Talagrand proved [34] the following important nonlinear Dvoretzky-like theorem,
where the notion of dimension of (X, d) is interpreted to be 2 (X, d). Theorem 1.10 below
is stated slightly differently in [34], but it easily follows from a combination of [34, Lem. 6],
[34, Thm. 11], and [34, Prop. 13].
Theorem 1.10 ([34]). There are universal constants c, D (0, ) such that every finite
metric space (X, d) has a subset S X that embeds into an ultrametric space with distortion
D and 2 (S, d) > c2 (X, d).
Theorem 1.10 is of major importance since it easily implies Talagrands majorizing mea-
sures theorem. Specifically, psuppose that {Gx }xX is a centered Gaussian process and for
x, y X we have d(x, y) = E[(Gx Gy )2 ]. Talagrands majorizing measures theorem as-
serts that E[supxX Gx ] > K2 (X, d), where K (0, ) is a universal constant. Let S X
be the subset obtained from an application of Theorem 1.10 to (X, d). Since ultrametric
spaces are isometric
p to subsets of Hilbert space, there is a Gaussian process {Hx }xS such
that (x, y) = E[(Hx Hy )2 ] is an ultrametric on S and d(x, y) 6 (x, y) 6 Dd(x, y) for
all x, y S. Hence 2 (S, ) > 2 (S, d) > c2 (X, d), and a standard application of Slepians

7
lemma (see [34, Prop. 5]) yields E[supxX Gx ] > E[supxS Gx ] > D1 E[supxS Hx ]. This
shows that due to Theorem 1.10 it suffices to prove the majorizing measures theorem when
(X, d) is an ultrametric space itself. Ultrametric spaces have a natural tree structure (based
on nested partitions into balls; see e.g. [5, Sec. 3.1]), and they can be embedded into Hilbert
space so that disjoint subtrees are orthogonal. For Gaussian processes orthogonality means
independence, which indicates how the ultrametric structure can be harnessed to yield a
direct and short proof of the majorizing measures theorem for ultrametrics 2. This striking
application of a metric Dvoretzky-type theorem is of great importance to several areas; we
refer to [35, 36] for an exposition of some of its many applications.
In a forthcoming paper [26] we show how Theorem 1.5 implies Talagrands nonlinear
Dvoretzky theorem. The deduction of Theorem 1.10 in [26] is based on the ideas presented
here, but it will be published elsewhere due to its length. Our proof of Theorem 1.5 does
not borrow from Talagrands proof of Theorem 1.10, and we do not see how to use Tala-
grands approach in order to deduce the general covering statement of Theorem 1.5. It is
an interesting open question to determine whether Talagrands method is relevant to the
setting of Theorem 1.5. Beyond being simpler, Talagrands original argument has addi-
tional advantages over our approach; specifically, it yields the important generic chaining
method [35, 36].
The nonlinear Dvoretzky theorems that are currently known, including variants of Theo-
rem 1.10 for other functionals that are defined similarly to 2 , differ from each other in the
notion of dimension, or largeness, of a metric space that they use. While we now have a
general nonlinear Dvoretzky theorem that contains the Bourgain-Figiel-Milman, Talagrand
and Tao phenomena as special cases, one might conceivably obtain a characterization of
notions of dimension of metric spaces for which a nonlinear Dvoretzky theorem can be
proved. We therefore end this introduction with an open-ended and purposefully somewhat
vague direction for future research.
Question 1.11. What are the notions of dimension of metric spaces that yield a nonlinear
Dvoretzky theorem in the sense that every (compact) metric space can be shown to contain
a subset of proportional dimension that well-embeds into an ultrametric space? At present
we know this for the following notions of dimension: log |X|, dimH (X), 2 (X) (and some
natural variants of these notions). Is there an overarching principle here?
2. Reduction to finite metric spaces
In this section we use a simple compactness argument to show that it suffices to prove
Theorem 1.5 and Theorem 1.9 when (X, d) is a finite metric space. Before doing so we fix
some standard terminology.
As we have already noted earlier, given a metric space (X, d) and r > 0, the closed ball
centered at x X of radius r is denoted B(x, r) = {y X : d(x, y) 6 r}. Open balls
are denoted by B (x, r) = {y X : d(x, y) < r}. We shall use this notation whenever the
metric space in question will be clear from the context of the discussion, but when we will
need to argue about several metrics at once we will add a subscript indicating the metric with
respect to which balls are taken. Thus, we will sometimes use the notation Bd (x, r), Bd (x, r).
Similar conventions hold for diameters of subsets of X: given a nonempty A X we denote
2According to Talagrand [36, Sec. 2.8], Fernique was the first to observe that the majorizing measures
theorem holds for ultrametrics. See [34, Prop. 13] for a short proof of this fact.

8
diam(A) = supx,yA d(x, y) whenever the underlying metric is clear from the context, and
otherwise we denote this quantity by diamd (A).
Given two nonempty subsets A, B X we denote as usual
d(A, B) = inf d(x, y), (6)
xA
yB

and we will also use the standard notation d(x, A) = d({x}, A). The Hausdorff distance
between A and B is denoted
 
dH (A, B) = max sup d(x, B), sup d(y, A) . (7)
xA yB

Lemma 2.1. Fix D, c > 1 and (0, 1]. Suppose that any finite metric measure space
(X, d, ) has a subset S X that embeds with distortion D into an ultrametric space, such
that every family of balls {B(xi , ri )}iI that covers S satisfies
X
(B(xi , cri )) > (X) . (8)
iI

Then any metric measure space (X, d, ) has a closed subset S X that embeds with distor-
tion D into an ultrametric space, such that every family of balls {B(xi , ri )}iI that covers S
satisfies (8).
Proof. Let (X, d, ) be a metric measure space and let Xn be a n1 -net in X, i.e., d(x, y) > n1
for all distinct x, y Xn , and d(x, Xn ) 6 n1 for all x X. Since X is compact, Xn is finite.
Write Xn = {xn1 , xn2 , . . . , xnkn } and for j {1, . . . , kn } define
def
jn (x) = min {i {1, . . . , kn } : d (x, xni ) = d(x, Xn )} .
Consider the Voronoi tessellation {V1n , . . . , Vknn } 2X given by
def
Vjn = {x X : jn (x) = j} .
Thus {V1n , . . . , Vknn } is a Borel partition of X, and we can define a measure n on Xn by
n (xnj ) = (Vj ). Note that by definition n (Xn ) = (X).
The assumption of Lemma 2.1 applied to the finite metric measure space (Xn , d, n ) yields
a subset Sn Xn , an ultrametric space (Un , n ) and a mapping fn : Sn Un such that
d(x, y) 6 n (fn (x), fn (y)) S 6 Dd(x, y) for all x, y Sn . Moreover, if {zj }jJ Xn and
{rj }jJ [0, ) satisfy jJ B(zj , rj ) Sn then
X
n (Xn B(zj , crj )) > n (Xn ) = (X) . (9)
jJ

Let U be a free ultrafilter over N. Since the Hausdorff metric dH (recall (7)) on the space
of closed subsets of X is compact (e.g. [10, Thm. 7.3.8]), there exists a closed subset S X
such that limnU dH (S, Sn ) = 0. Q
Define : S S [0, ) as follows. For x, y S there are (xn )
n=1 , (yn )n=1 n=1 Sn
such that limnU d(x, xn ) = limnU d(y, yn ) = 0. Set (x, y) = limnU n (fn (xn ), fn (yn )).
This is well defined, i.e., (x, y) does not depend on the choice of (xn )
n=1 , (yn )n=1 . Indeed,

9
Q
if (x0n ) 0
n=1 , (yn )n=1 n=1 Sn also satisfy limnU d(x, x0n ) = limnU d(y, yn0 ) = 0 then
lim n (fn (xn ), fn (yn ))
nU
6 lim n (fn (x0n ), fn (yn0 )) + lim n (fn (xn ), fn (x0n )) + lim n (fn (yn ), fn (yn0 ))
nU nU nU
6 lim n (fn (x0n ), fn (yn0 )) + D lim d(xn , x0n ) + D lim d(yn , yn0 )
nU nU nU
= lim n (fn (x0n ), fn (yn0 )),
nU

so that by symmetry limnU n (fn (xn ), fn (yn )) = limnU n (fn (x0n ), fn (yn0 )). It is immediate
to check that d(x, y) 6 (x, y) 6 Dd(x, y) for all x, y S S and that is an ultrametric on S.
Now, let {xi }i=1 X and {ri }i=1 [0, ) satisfy

i=1 B(xi , ri ) S. Fix > 0. For
every i N there is i (0, 1) such that
 1/
(B(xi , cri + ci )) 6 (B(xi , cri )) + i . (10)
2
S
Since S is compact, there exists a finite subset I N such that iI B(xi , ri + i /2) S.
Denote = miniI i . By definition of S there exists n N such that n > 8/ and
, S) < /8. For every i I let zin Xn satisfy d(zin , xi ) S
dH (SnS = d(xi , Xn ) 6 1/n < /8.
Now, iI B(zin , ri + 3i /4) Sn because dH (Sn , S) < /8 and iI B(xi , ri + i /2) S.
An application of (9) now yields the bound
X   
n 3ci
n Xn B zi , cri + > (X) . (11)
iI
4

By the definition of n ,
  
n 3ci
n Xn B zi , cri +
4
[   
n n n 3ci
= Vj : j {1, . . . , kn } xj B zi , cri +
4
  
3ci 1
6 B zin , cri + +
4 n
  
3ci 2
6 B xi , cri + +
4 n
6 (B (xi , cri + ci ))
(10)  1/
6 (B(xi , cri )) + i . (12)
2
Hence,
(11)(12) X
  1/ 

(X) 6 (B(xi , cri )) + i
iI
2


X  X
6 (B(xi , cri )) + 6 + (B(xi , cri )) . (13)
iI
2i i=1

Since (13) holds for all > 0, the proof of Lemma 2.1 is complete. 

10
Assumptions. Due to Lemma 2.1 we assume from here through the end of Section 9 that
(X, d, ) is a finite metric measure space. By restricting to the support of , we assume
throughout that ({x}) > 0 for all x X. By rescaling the metric, assume also that
diam(X) = 1.
3. Combinatorial trees and fragmentation maps
The ensuing arguments rely on a variety of constructions involving combinatorial trees.
We will work only with finite rooted trees, i.e., finite graph-theoretical trees T with a distin-
guished vertex r(T ) called the root of T . We will slightly abuse notation by identifying T
with its vertex set, i.e., when we write v T we mean that v is a vertex of T . We shall say
that u T is an ancestor of v T r {u} if u lies on the path joining v and r(T ). In this case
we also say that v is a descendant of u. We say that v is a weak descendant (respectively
weak ancestor) of u if it is either a descendant (respectively ancestor) of u or v = u. If u is
either a weak ancestor of v or a weak descendant of v we say that u and v are comparable,
and otherwise we say that they are incomparable. The leaves of T , denoted L(T ) T , is
the set of vertices of T that do not have descendants.
Definition 3.1 (Cut set). Let T be a rooted tree. A subset S T is called a cut set of T
if any root-leaf path in T intersects S. Equivalently, S is a cut set of T if every u T is
comparable to a vertex in S. See [29, Ch. 4 & Sec. 12.4].
If v T r {r(T )} then we denote by p(v) = pT (v) its parent in T , i.e., the vertex adjacent
to v on the path joining v and r(T ). We say that v T r {r(t)} is a child of u T if
p(v) = u, and the set p1 (u) = {v T : p(v) = u} is the set of children of u. Thus
L(T ) = {u T : p1 (u) = }. If u, v T r {r(T )} are distinct and satisfy p(u) = p(v)
then we say that u and v are siblings in T .
The depth of u T , denoted depthT (u), is the number of edges on the path joining
u and r(T ). Thus depthT (r(T )) = 0. The least common ancestor of u, v T , denoted
lca(u, v) = lcaT (u, v), is the vertex of maximal depth that is an ancestor of both u and v.
Definition 3.2 (Subtree). Let T be a finite rooted tree. A subtree T 0 of T is a connected
rooted subgraph of T whose set of leaves is a subset of the leaves of T , i.e., L(T 0 ) L(T ).
Given u T , we denote by Tu T the subtree rooted at u, i.e., the tree consisting of all
the weak descendants of u in T , with the edges inherited from T . Thus r(Tu ) = u.
Definition 3.3. Let T be a rooted tree and A T . For u T define DT (u, A) T to be
the set of all v A such that v is a descendant of u and no ancestor of v is also in A and is
a descendant of u. Note that DT (u, A) = if u has no descendants in A, and DT (u, A) is a
cut set of the subtree Tu if A Tu is a cut-set in Tu (this happens in particular if A contains
the leaves of Tu ). We also define

DT (u, A) if u T r A,
DT (u, A) = (14)
{u} if u A.

Trees interact with metric spaces via the notion of fragmentation maps.
Definition 3.4 (Fragmentation map). Let (X, d) be a finite metric space. A fragmentation
map of X is a function F : T 2X , where T is a finite rooted tree that satisfies the following
conditions.

11
F(r(T )) = X.
If v L(T ) is a leaf of T then F(v) is a singleton, i.e., F(v) = {x} for some x X.
If v T r {r(T )} the F(v) F(p(v)).
If u, v T are incomparable then F(u) F(v) = .
In what follows, given a fragmentation map F : T 2X , we will use the notation Fu = F(u).

Definition 3.5 (Boundary of a fragmentation map). The boundary of a fragmentation map


F : T 2X is a new map F : T 2X defined as follows. For u T the set F(u) = Fu
is the subset of X corresponding to the image under F of the leaves of the subtree Tu , i.e.,
[
Fu = Fv .
vL(Tu )

Note that we always have Fu Fu .

Definition 3.6 (Partition map). A partition map is a fragmentation map F : T 2X such


that Fr(T ) = X. Note that in this case F = F.

Up to this point the metric on X did not play any role. The following definition is one out
of two definitions that tie the structure of a fragmentation map F : T 2X to the geometry
of X (the second definition, called the separation property, will be introduced in Section 5).

Definition 3.7 (Lacunary fragmentation map). Given K, (0, ), a fragmentation map


F : T 2X is (K, )-lacunary if for every q T and every u T such that u is a weak
descendant of q and u has at least two children, i.e., |p1 (u)| > 1, we have

diam (Fq ) 6 K depthT (u)depthT (q) min d (Fv , Fw ) . (15)


v,wp1 (u)
v6=w

Lemma 3.8. Let F : T X be a (K, )-lacunary fragmentation map of a metric space


(X, d). Then (Fr(T ) , d) embeds with distortion K into an ultrametric space.

Proof. For x, y Fr(T ) there are a, b L(T ) such that Fa = {x} and Fb = {y}. Define

(x, y) = diam Flca(a,b) .

Since x, y Flca(a,b) we have d(x, y) 6 (x, y). Assume that x 6= y. Then lca(a, b) has distinct
children v, w p1 (lca(a, b)) such that x Fv and y Fw . An application of (15) to
q = u = lca(a, b) shows that (x, y) 6 Kd(x, y). It remains to note that is an ultrametric.
Indeed, take a, b, c L(T ) and write Fa = {x}, Fb = {y}, Fc = {w}. If lca(a, b) is a weak
descendant of lca(b, c) then Flca(a,b) Flca(b,c) , implying that (x, y) 6 (y, z). Otherwise
lca(a, b) {lca(a, c), lca(b, c)}, implying that (x, y) = max{(x, z), (y, z)}. 

The proof of Lemma 3.8 did not use the full strength of Definition 3.7. Specifically, the
parameter did not appear, and we could have used a weaker variant of (15) in which the
left hand side is diam(Fq ) instead of diam(Fq ). The full strength of the (K, )-lacunary
condition will be used in the ensuing arguments since they allow us to have better control
on restrictions of fragmentation maps to subtrees of T .

12
4. From fragmentation maps to covering theorems
Here we show how a lacunary fragmentation map which satisfies a certain cut-set inequality
can be used to prove a covering theorem in the spirit of the conclusion of Theorem 1.5. This
is the content of the following lemma.
Lemma 4.1. Fix K, (0, ) and (0, 1). Let (X, d, ) be a finite metric measure
space. Assume that there exists a (K, )-lacunary fragmentation map G : T 2X such that
every leaf ` L(T ) has no siblings, and furthermore Gp(`) = G` . Suppose also that for any
cut-set G of T we have
X
(Gp(v) ) > (X) ,
vG

where if v T is the root then we set p(v) = v.


Then for any {xi }iI X and {ri }iI [0, ) such that the d-balls {Bd (xi , ri )}iI cover
Gr(T ) , we have
X  
Bd xi , 1 + 2K 2 ri > (X) . (16)
iI

Proof. Without loss of generality assume that Gr(T ) Bd (xi , ri ) 6= for all i I. Let be
the ultrametric induced by G on Gr(T ) , as constructed
 in the proof of Lemma 3.8. Thus
for x, y Gr(T ) we have (x, y) = diamd Glca(a,b) , where a, b L(T ) satisfy Ga = {x} and
Gb = {y}. Note that this definition implies that
v T, diam (Gv ) = diamd (Gv ). (17)
By Lemma 3.8 we know that
x, y Gr(T ) , d(x, y) 6 (x, y) 6 Kd(x, y). (18)
For every i I choose yi Gr(T ) satisfying
d(xi , yi ) = min d(xi , y) 6 ri . (19)
yGr(T )

Then Bd (yi , 2ri ) Bd (xi , ri ). Hence the balls {Bd (yi , 2ri )}iI also cover Gr(T ) . By (18) we
have B (yi , 2Kri ) Bd (yi , 2ri ) Gr(T ) , so we also know that the -balls {B (yi , 2Kri )}iI
cover Gr(T ) .
For i I choose vi T as follows. If B (yi , 2Kri ) is a singleton then vi is defined to be
the leaf of T such that Gvi = {yi }. Otherwise pick vi to be the highest ancestor of yi in T
such that
diam (Gvi ) 6 2Kri (20)
and vi has at least twoSchildren. Then B (yi , 2Kri ) = Gvi . Hence, since {B (yi , 2Kri )}iI
cover Gr(T ) , we have iI L (Tvi ) = L(T ). If we set G = {vi }iI then we conclude that G is
a cut-set of T . Our assumption therefore implies that
X 
Gp(vi ) > (X) . (21)
iI

When vi has at least two children we deduce from the fact that G is (K, )-lacunary that

diamd Gp(vi ) 6 K diamd (Gvi ). (22)

13
(Recall Definition 3.7 with q = p(vi ) and u = vi .) When vi is a leaf our assumptions imply
that Gvi and Gp(vi ) are both singletons, and therefore diamd Gp(vi ) = diamd (Gvi ) = 0,
so (22) holds in this case as well. Hence for every i I and z Gp(vi ) we have
(19) () 
d(z, xi ) 6 d(xi , yi ) + d(z, yi ) 6 ri + d(z, yi ) 6 ri + diamd Gp(vi )
(22) (18) (20)
6 ri + K diamd (Gvi ) 6 ri + K diam (Gvi ) 6 ri + 2K 2 ri , (23)
where () follows from the fact that yi Gvi Gvi Gp(vi ) . The validity of (23) for
all z Gp(vi ) is the same as the inclusion Gp(vi ) Bd (xi , (1 + 2K 2 ) ri ). Now (16) follows
from (21). 
In light of Lemma 4.1, our goal is to construct a fragmentation map F : T 2X satisfying
the assumptions of Lemma 4.1 with = 1, such that (Fr(T ) , d) embeds into an ultrametric
space with distortion O(1/). Note that the (K, )-lacunary assumption in Lemma 4.1
implies by Lemma 3.8 that (Fr(T ) , d) embeds into an ultrametric space with distortion K.
However, more work will be needed in order to obtain the desired O(1/) distortion.
In what follows we use the following notation.
Definition 4.2. Given D (2, ) let (D) (0, 1) denote the unique solution of the
equation
2
= (1 ) 1 . (24)
D
It is elementary to check that
2e
D (2, ), (D) > 1 , (25)
D
and
c
(0, 1/2), (2 + ) > , (26)
log(1/)
where c (0, ) is a universal constant.
The following key lemma describes the fragmentation map that we will construct.
D2

Lemma 4.3. Fix D (2, ), an integer k > 2, and 0, 3D+2 . Let (X, d, ) be a finite
metric measure space of diameter 1. Then there exists a fragmentation map G : T 2X
with the following properties.
Every leaf ` L(T ) has
 no siblings, and furthermore GpT (`) = G` .
2 2
G is 2
13
4k , 4k
-lacunary.

Gr(T ) , d embeds into an ultrametric space with distortion D.
Every cut-set G T (recall Definition 3.1) satisfies
(1 k1 )2 ( 13 D) 1 2 13
> (X)(1 k ) ( 1+ D) ,
X
GpT (v) 1+
(27)
vG

where pT (v) is the parent of v in T if v is not the root, and the root if v is the root.
Lemma 4.3 will be proved in Section 5. Assuming its validity for the moment, we now
proceed to use it in combination with Lemma 4.1 to prove Theorem 1.5 and Theorem 1.9.

14
Proof of Theorem 1.5. By Lemma 2.1 we may assume that (X, d, ) is a finite metric measure
2
space. Fix an integer 10

6 k 6 11
and set = 201
. Then (1 ) 1 k1 (0, 1), so we can
define ! !
1+ 1 11 1 1
D= 1 2 = 2 . (28)
1 3 1 1 7 1 1
k k
Equivalently,
 2  
1 1 3
1 D = 1 .
k 1+
Due to (24), for every s (0, 1) we have 1 (s) = 2(1 s)1 ss/(1s) > 2. Hence it follows
from (28) that D > 2(1 + )/(1 3 ), or equivalently < (D 2)/(3D + 2). By (25) we
have 1 (s) 6 2e/(1 s). Therefore,
11 2e 42e(10 )2 9
D6 1 = 6 , (29)
7 1 (1/10)2 17(8 + )
where the last inequality in (29) is elementary. The required conclusion now follows from
2 2
Lemma 4.3 and Lemma 4.1. Note that we get the bound c = O(k ) = eO(1/ ) . 
Proof of Theorem 1.9. Again, using Lemma 2.1 we may assume that (X, d, ) is a finite
metric measure space. Apply Lemma 4.3 with D = 2 + , k = 2 and = /9. Denote the
exponent in (27) by s = 12 ((9 3)(2 + )/(9 + )). By (26) there is a universal constant
c (0, ) such that s > t, where t = c/ log(1/). Let G : T 2X be the fragmentation
obtained from Lemma 4.3, and let G be a cut-set in T . Then by (27) we have,
!1/t !1/s
X t X s
GpT (v) > GpT (v) > (X).
vG vG

We can therefore apply Lemma 4.1 with = t, K = 2 16 /(1 3 ) and = 16 , obtaining


Theorem 1.9. Note that this shows that c0 can be taken to be a constant multiple of 16 . 

5. Asymptotically optimal fragmentation maps: proof of Lemma 4.3


It remains to prove Lemma 4.3 in order to establish Theorem 1.5 and Theorem 1.9. The
proof of Lemma 4.3 decomposes naturally into two parts. The first part yields a fragmen-
tation map F : T 2X that satisfies the desired cut-set inequality (27), but the distortion
of (Fr(T ) , d) in an ultrametric space is not good enough. The second part improves the
embeddability of (Fr(T ) , d) into an ultrametric space by performing further pruning.
We begin with the second part since it is shorter and simpler to describe. In order to be
able to improve the embeddability of (Fr(T ) , d) into an ultrametric space, we will use the
following property.
Definition 5.1 (Separated fragmentation map). Given > 0 and a fragmentation
 map
X
F : T 2 , a vertex u T is called -separated if for every x Fr(T ) r (Fu ) we have
d (x, Fu ) > diam (Fu ) . (30)
The map F : T 2X is called -separated if all the vertices u T are -separated.

15
The following very simple lemma exploits the fact that the class of ultrametrics is closed
under truncation. This fact will serve as a useful normalization in the ensuing arguments.
Lemma 5.2. Let (X, d) be a bounded metric space that embeds with distortion D into an
ultrametric space. Then there exists an ultrametric on X satisfying
d(x, y) 6 (x, y) 6 Dd(x, y) for all x, y X,
diamd (X) = diam (X).
Proof. We are assuming that there exist A, B > 0 and an ultrametric 0 on X that satisfies
Ad(x, y) 6 0 (x, y) 6 Bd(x, y) for all x, y X, where B/A 6 D. We can therefore define
= min{0 /A, diamd (X)}. 
The last ingredient that we need before we can state and prove the lemma that describes
how to improve the embeddability of (Fr(T ) , d) into an ultrametric space is a weighted
version of nonlinear Dvoretzky theorem for finite metric spaces. As discussed in the in-
troduction, it was proved in [30] that for every D > 1, every n-point metric space (X, d)
contains a subset of size n(D) that embed in ultrametric with distortion at most D, where
(D) is defined in (24). We will need the following generalization of this result.
Theorem 5.3. For every D > 2, every finite metric space (X, d) and every w : X (0, ),
there exists S X that embeds with distortion D into an ultrametric space and satisfies,
!(D)
X X
w(x)(D) > w(x) . (31)
xS xX

With some minor changes, the proof in [30] also applies to the more general weighted
setting of Theorem 5.3. We prove Theorem 5.3 in Section 9 by sketching the necessary
changes to the argument in [30].
Assuming the validity of Theorem 5.3, we are now ready to improve the ultrametric
distortion of a fragmentation map by performing additional pruning. We use the metric
composition technique of [5], which takes a vertex and its children in the tree associated to
the fragmentation map, deletes some of these children, and arranges the remaining children
into a new tree structure. The deletion is done by solving a nonlinear Dvoretzky problem
for weighted finite metric spaces, i.e., by applying Theorem 5.3.
Lemma 5.4. Fix D (2, ) and (0, ). Let (X, d) be a finite metric space. Suppose
that F : T 2X is a fragmentation map which is -separated. Suppose also that there is
a weight function w : T (0, ) which is subadditive, i.e., that for every non-leaf vertex
u T r L(T ), X
w(v) > w(u), (32)
vp1
T (u)

Then there exists a subtree T 0 of T with the same root such that the restricted fragmentation
map G = F|T 0 satisfies the following properties.
 
Gr(T 0 ) , d embeds into an ultrametric space with distortion D 1 + 2 .


Every non-leaf vertex u T 0 r L(T 0 ) satisfies


X
w(v)(D) > w(u)(D) . (33)
vp1
T0
(u)

16
Proof. Before delving into the details of proof, the reader may want to consult Figure 1, in
which the strategy of the proof is illustrated.

Figure 1. A schematic illustration of the proof of Lemma 5.4. The


first figure from the left depicts three levels of the fragmentation map,
the middle level being separated. In the second figure from the left
we consider a certain induced metric (see (34)) on the clusters in the
middle level. Due to the separation property, this metric approximates
the actual distances between points in different clusters. In the third
figure from the left we have applied the weighted finite Dvoretzky the-
orem, i.e., Theorem 5.3, to the middle level clusters, thus obtaining an
appropriately large subset of clusters on which the induced metric is
approximately an ultrametric. The rightmost figure describes the tree
representation of this new ultrametric.

For every vertex u T r L(T ) let Cu = p1


T (u) be the set of children of u in T . Let du be
a metric defined on Cu as follows

diamd ((Fx ) (Fy )) if x 6= y,
du (x, y) = (34)
0 if x = y.
The validity of the triangle inequality for du is immediate to verify. By the definition of (D)
there exists a subset Su Cu such that
!(D)
X (31) X (32)
w(x)(D) > w(x) > w(u)(D) . (35)
xSu xCu

and (Su , du ) embeds with distortion D into an ultrametric space. By Lemma 5.2 there exists
an ultrametric u on Su such that every x, y Su satisfy

du (x, y) 6 u (x, y) 6 min diamdu (Su ), Ddu (x, y)
( ! )
(34) [
6 min diamd Fx , Ddu (x, y) . (36)
xSu

The subtree T 0 T is now defined inductively in a top-down fashion as follows: declare


r(T ) T 0 and if u T is a non-leaf vertex that was already declared to be in T 0 , add the
vertices in Su to T 0 as well. Inequality (33) follows from (35). It remains to prove that
Gr(T 0 ) , d embeds into an ultrametric space with distortion D (1 + 2/). To this end fix
p, q Gr(T 0 ) and choose the corresponding a, b L(T 0 ) such that Ga = {p} and Gb = {q}.
Let u = lcaT 0 (a, b) = lcaT (a, b) and choose x, y Su that are weak ancestors of a and b,

17
respectively. Define (p, q) = u (x, y). Now,
(34) (36)
d(p, q) = d(Ga , Gb ) 6 diamd ((Fx ) (Fy )) = du (x, y) 6 u (x, y) = (p, q).
The corresponding lower bound on d(p, q) is proved as follows, using the assumption that
the fragmentation map F is -separated.
 
(p, q) (36) (34) (30) 2
6 du (x, y) 6 diamd (Fx ) + diamd (Fy ) + d (Fx , Fy ) 6 1 + d(p, q).
D
We now argue that is an ultrametric on Gr(T 0 ) . This is where we will use the truncation

S
in (36), i.e., that for all u T r L(T ) we have diamu (Su ) 6 diamd xSu Fx . Take
p1 , p2 , p3 Gr(T 0 ) and choose the corresponding a1 , a2 , a3 L(T 0 ) such that Gai = {pi } for
i {1, 2, 3}. By relabeling the points if necessary, we may assume that u = lcaT (a1 , a2 ) is a
weak descendant of v = lcaT (a2 , a3 ). If u = v take x1 , x2 , x3 Su that are weak ancestors of
a1 , a2 , a3 , respectively. Since u is an ultrametric, it follows that
(p1 , p2 ) = u (x1 , x2 ) 6 max {u (x1 , x3 ), u (x3 , x2 )} = max {u (p1 , p3 ), u (p3 , p2 )} .
If, on the other hand, u is a proper descendant of v then choose x1 , x2 Su that are
weak ancestors of a1 , a2 (respectively), and choose s, t Sv that are weak ancestors of u, a3
(respectively). Then,

!
(36) [
(p1 , p2 ) = u (x, y) 6 diamd Fw 6 diamd (Fs )
wSu

(34) (36)
6 diamd ((Fs ) (Ft )) = dv (s, t) 6 v (s, t) = (p1 , p3 ) = (p2 , p3 ).
This establishes the ultratriangle inequality for , completing the proof of Lemma 5.4. 
The next lemma establishes the existence of an intermediate fragmentation map with
useful geometric properties; its proof is deferred to Section 6.
Lemma 5.5. Fix (0, 1/3) and integers m, h, k > 2 with h > 2k 2 . Let (X, d, ) be a finite
metric measure space of diameter 1. Then there exists a fragmentation map F : T 2X
with the following properties.
{1} All the leaves of the tree T are at depth mh.
{2} For every u T we have
diam(Fu ) 6 depthT (u) . (37)
{3} Denote by R T the set of vertices at depths which are integer multiples of h. Then
for every non-leaf u R,
1 2 1 2
(Fv )(1 k ) > (Fu )(1 k ) .
X
(38)
vDT (u,R)

Recall that DT (, ) is given in Definition 3.3.


{4} There is a subset S T containing the root of T such that R and S are alternating in
the following sense. For every u, v R such that depthT (v) = depthT (u) + h and v is
a descendant of u, there is one and only one w S such that w lies on the path joining
u and v and depthT (u) < depthT (w) 6 depthT (v).

18
13
{5} The vertices of S are 2
-separated (recall Definition 5.1).
2 2h 1

{6} F is 13 , -lacunary (recall Definition 3.7).
The vertices of the subset R T of Lemma 5.5 satisfy an inductive inequality (38) on the
measures of their images that will allow us to (eventually) deduce the covering property (27)
of Lemma 4.3. Figure 2 contains a schematic depiction of the fact that the levels of R and
S alternate.

R, S 0

S
R h

S
R 2h
S

R 3h

S
R 4h

Figure 2. A schematic depiction of the tree T corresponding to the


fragmentation map F of Lemma 5.5. The vertices of R are on the
dotted lines. The vertices of S are on the curved solid lines. On every
root-leaf path in T the vertices in R and S alternate.

We are now in position to prove Lemma 4.3 using Lemma 5.4 and assuming the validity
of Lemma 5.5 (recall that Lemma 5.5 will be proved in Section 6).
Proof of Lemma 4.3. Let k, be as in Lemma 4.3. Denote h = 2k 2 and fix m N satisfying
min d(x, y) > (m2)h+1 . (39)
x,yX
x6=y

Apply Lemma 5.5 with the parameters , m, h, k as above, obtaining a fragmentation map
F : T 1 2X with corresponding subsets S, R T 1 . Let T 2 be the tree induced by T 1
on S, i.e., join u, v S by an edge of T 2 if u is an ancestor of v in T 1 and any w S
that is an ancestor of v in T 1 is a weak ancestor of u. This is the same as the requirement
v DT 1 (u, S). Let S : T 2 2X be the fragmentation map obtained by restricting F to S.
To check that S is indeed a fragmentation map we need to verify that if u L(T 2 ) then
Fu is a singleton. To see this let v = pT 2 (u) be the parent in T 2 of u. Since all the leaves
of T are at depth mh, we have depthT 1 (v) > (m 2)h + 1. Using (37) we deduce that
diam(Fu ) 6 diam(Fv ) 6 (m2)h+1 , implying that Su and Sv are both singletons due to (39).
Since by Lemma 5.5 we know that the vertices in S are 13 2
-separated in the fragmentation
13
map F, it follows that the fragmentation map S is 2
-separated. Lemma  5.5 also ensures
2 2h 1 2 2h 2h

that F is 13 , -lacunary. This implies that S is 13 , -lacunary. Indeed,
due to Lemma 5.5 we know that if u S and v DT 1 (u, S) is a child of u in T 2 then
depthT 1 (v) 6 depthT 1 (u) + 2h 1. This implies that if q, u S are such that u is a

19
weak descendant of q in T 2 then depthT 1 (u) depthT 1 (q) 6 2h (depthT 2 (u) depthT 2 (q)).
Hence, if v, w DT 1 (u, S) are distinct children of u in T 2 , choose distinct x, y T 1 that
are children ofu in T 1 and weak ancestors of v, w (respectively), and use the fact that F is
2
13
2h , 1 -lacunary to deduce that

2 2h
diamd (Sq ) = diamd (Fq ) 6 (depthT 1 (u)depthT 1 (q)) d (Fx , Fy )
1 3
2 2h
6 2h(depthT 2 (u)depthT 2 (q)) d (Sx , Sy ) .
1 3
Define wR : R (0, ) by top-down induction as follows. Set
1 2
wR (r) = (X)(1 k ) , (40)
1
where r is the root of T . If u R is not a leaf and v DT 1 (u, R) then define
wR (u) 1 2
wR (v) = P 2 (Fv )(1 k ) . (41)
1 k1
zDT 1 (u,R)
(Fz )( )

Thus for every non-leaf u R we have


X
wR (u) = wR (v). (42)
vDT 1 (u,R)

Moreover, it follows from the recursive definition (41) combined with (38) that
1 2
u R, wR (u) 6 (Fu )(1 k ) . (43)
Recalling the notation DT (x, A) as given in (14), by summing (42) we see that for all
u S r L(T 2 ) we have
X X X
wR (y) = wR (x). (44)
xD 1 (u,R) yDT 1 (x,R) xD 1 (u,R)
T T

Notice that [ [
DT 1 (x, R) = DT 1 (v, R), (45)
xD 1 (u,R) vDT 1 (u,S)
T

where the unions on both sides of (45) are disjoint. Hence,


X (44)(45) X X
wR (x) = wR (z). (46)
xD 1 (u,R) vDT 1 (u,S) zD 1 (v,R)
T T

Define wS : S (0, ) by
X
wS (u) = wR (z). (47)
zD 1 (u,R)
T

Then for all u S r L(T 2 ) we have


X X (46)(47) X (47)
wS (v) = wS (v) = wR (x) = wS (u).
vp1 vDT 1 (u,S) xD 1 (u,R)
T2 (u) T

20
This establishes condition (32) of Lemma 5.4 for the weighting wS of T 2 . Before applying
Lemma 5.4 we record one more useful fact about wS . Recall that for u S the vertex pT 2 (u)
is r if u = r, and otherwise it is the first proper ancestor of u in T 1 which is in S. Take
u0 DT 1 (pT 2 (u), R) which is a weak ancestor of u in T 1 . Then FpT 2 (u) Fu0 , and therefore

(1 k1 )2 1 2 (43)
FpT 2 (u) > (Fu0 )(1 k ) > wR (u0 )
(42) X X (47)
= wR (x) = wR (x) = wS (u). (48)
xDT 1 (u0 ,R) xD 1 (u,R)
T

2 X 2
Apply Lemma 5.4 to S : T 2 and wS : T (0, ), with = (1 3 )/(2 ) and the
parameter D of Lemma 5.4 replaced by D/(1 + 2/) = D(1 3 )/(1 + ). Note that our
assumption < (D 2)/(3D + 2) guarantees that this new value of D is bigger than 2, so we
are indeed allowed to use Lemma 5.4. We therefore obtain a subtree T T 2 with the same
root, such that the restricted fragmentation map G = S|T satisfies the following properties.
(Gr(T ) , d) embeds into an ultrametric space with distortion D.
Every non-leaf vertex u T r L(T ) satisfies
13 (33) 13
wS (v)( 1+ D) > wS (u)( 1+ D) .
X
(49)
vp1
T (u)

Let G T be a cut-set of T . Define G0 to be a subset of G which is still a cut-set


and is minimal with respect to inclusion. Assume inductively that we defined a cut-set Gi
of T which is minimal with respect to inclusion. Let v Gi be such that depthT (v) is
maximal. Let u = pT (v). By the maximality of depthT (v), since Gi is a minimal cut-set of
T we necessarily have p1 T (u) Gi , i.e., all the siblings of v in T are also in Gi . Note that
1
Gi = (Gi {u}) r pT (u) is also a cut-set of T , so let Gi+1 be a subset of G0i which is still a
0

cut-set of T and is minimal with respect to inclusion. Then,


13 13 13
wS (v)( 1+ D) = wS (v)( 1+ D) + wS (v)( 1+ D)
X X X

vGi vGi rp1


T (u) vp1
T (u)

(49) 13 13
wS (v)( 1+ D) > wS (v)( 1+ D) . (50)
X X
>
vG0i vGi+1

After finitely many iterations of the above process we will arrive at Gj = {r}. By concate-
nating the inequalities (50) we see that
13 13 13 (40)(47) 1 2 13
wS (v)( 1+ D) > wS (v)( 1+ D) > wS (r)( 1+ D) = (X)(1 k ) ( 1+ D) . (51)
X X

vG vG0

The
 desired inequality
 (27) follows from (51) and (48). Since S is 13
2
-separated and
2 2
2
13
4k , 4k -lacunary (recall that h = 2k 2 ), the same holds true for G since it is
the restriction of S to the subtree of T 2 . 
O(1/2 )
Remark 5.6. In Theorem 1.5, if one is willing to settle for ultrametric distortion e ,
instead of the asymptotically optimal O(1/) distortion, then it is possible to simplify

21
Lemma 4.3 and its proof. In particular, there is no need to apply Lemma 5.4, and con-
sequently also Theorem 5.3. Thus one can use the fragmentation map S introduced in
the
 proof of Lemma  4.3 as the fragmentation map produced by Lemma 4.3. Since S is
2 4k2 4k2
13
, -lacunary, Lemma 3.8 implies that (Sr(S) , d) embeds in an ultrametric
2 2
space with distortion 132
4k = eO(1/ ) . It is possible to further simplify the proof of the
cut-set inequality (27) in the proof of Lemma 4.3 by considering a different fragmentation
map R instead of S, defined as follows. Consider the tree T 3 induced by T 1 on R, and the
3 X 3
 map R : T
fragmentation  2 obtained by restricting F to T . Like S, the fragmentation
2 2
map R is 132
4k , 4k -lacunary, and the proof of (27) for R can now be performed by
only using the weight function wR , without the need to consider wS . Unlike S, the fragmen-
tation map R is not separated, and therefore cannot be used with Lemma 5.4. However, for
the above simplified argument, Lemma 5.4 and the separation property are not needed.

6. An intermediate fragmentation map: proof of Lemma 5.5


Here we prove Lemma 5.5. The proof uses two building blocks: Lemma 6.2, which con-
structs an initial partition map, and Lemma 6.5, which prunes a given weighted rooted tree.
The basic idea of the proof Lemma 5.5 can be described as follows. Lemma 6.2 constructs an
initial partition map together with a designated child for every non leaf vertex. The desig-
nated children have, roughly speaking, the largest weight among their siblings, and they are
also pairwise separated. The pruning step of Lemma 6.5 can now focus on the combinatorial
structure of the partition map, pruning the associated tree so as to keep at some levels only
the designated children of the level above. This guarantees the separation property as well
as the desired estimate (38).
The exact notion of size used to choose designated children is tailored to be compatible
with the ensuing pruning step, and is the content of the following definition. Observe that
any fragmentation map F : T 2X induces a weighting w : T (0, ) of the vertices of T
given by w(u) = (Fu ). For our purpose, we will need a modified version of it, described in
the following definition.
Definition 6.1 (Modified weight function). Fix integers h, k > 2 and let T be a finite
graph-theoretical rooted tree, all of whose leaves are at the same depth, which is divisible
by h. Assume that we are given w : T (0, ). Define a new function whk : T (0, ) as
follows. If u L(T ) then
k1
whk (u) = w(u) k .
Continue defining whk (u) by reverse induction on depthT (u) as follows.
 k1
k w(u) k if h | depthT (u),
wh (u) = P k (52)
vp1 (u) wh (v) if h - depthT (u).

Equivalently, if u T and (j 1)h < depthT (u) 6 jh for some integer j then
k1
X
whk (u) = w(v) k . (53)
vTu
depthT (v)=jh

22
Lemma 6.2. Let (X, d, ) be a finite metric measure space of diameter 1 and (0, 1/3).
For every triple of integers m, h, k > 2 there exists a fragmentation map F : T 2X with
the following properties.
All the leaves of the tree T are at depth mh.
F is a partition map, i.e., Fr(T ) = X.
For every u T we have
diam(Fu ) 6 depthT (u) . (54)
1
Every non-leaf vertex u T r L(T ) has a designated child c(u) p (u) such that
whk (c(u)) = max
1
whk (v) , (55)
vp (u)

where w : T (0, ) is given by w(u) = (Fu ) and whk : T (0, ) is the modified
weight function from Definition 6.1.
Suppose that u, v T r L(T ) satisfy depthT (u) = depthT (v) and u 6= v. Then
 1 3 depth (u)
d Fc(u) , Fc(v) > T . (56)
2
A schematic description of the partition map that is constructed in Lemma 6.2 is depicted
in Figure 3. Lemma 6.2 will be proved in Section 7.

Figure 3. A schematic depiction of two levels in the initial partition


map that is constructed in Lemma 6.2. In the lower level the darkness
of the cluster represent their weight whk ; darker means larger weight.
A thick line represents the designated child of the higher level cluster.
Notice that the designated child is the child of its parent of largest
weight, and that the designated children are far from each other.

Lemma 6.5 below is the pruning step. The appropriate setting for the pruning is a certain
class of weighted trees, which we now introduce; the definition below contains the tree of
Lemma 6.2 as a special case.
Definition 6.3 (Subadditive weighted tree with designated children). Fix integers h, k > 2.
A subadditive weighted tree with designated children is a triple (T, w, c) consisting of a finite
rooted graph-theoretical tree T , a mapping w : T (0, ) and for every non-leaf vertex
u T r L(T ) a designated child c(u) p1 (u), such that the following conditions hold
true.
All the leaves of T are at the same depth, which is divisible by h.

23
For every non-leaf vertex u T r L(T ),
X
w(u) 6 w(v). (57)
vp1 (u)

For every non-leaf vertex u T r L(T ),


whk (c(u)) = max
1
whk (v), (58)
vp (u)

where whk : T (0, ) is the modified weight function of Definition 6.1.


Definition 6.4 (Subtree sparsified at a subset). Fix two integers h, k > 2 and let let (T, w, c)
be a subadditive weighted tree with designated children (recall Definition 6.3). Let T 0 be a
subtree of T (see Definition 3.2) and S T 0 . We say that the subtree T 0 is sparsified at S if
every v S is a designated child and has no siblings in T 0 . For the purpose of this definition
we declare the root r(T ) to be a designated child, i.e., we allow r(T ) S. Thus v T is a
designated child if it is either the root of T or c(p(v)) = v.
Lemma 6.5. Fix two integers h, k > 2 with h > 2k 2 . Let (T, w, c) be a subadditive weighted
tree with designated children as in Definition 6.3 (thus all the leaves of T are at the same
depth, which is divisible by h, and the designated child map c satisfies (58)). Then there
exists a subtree T 0 of T with the same root as T , and two subsets R, S T 0 , both containing
the root of T 0 , with the following properties:
For any non-leaf u T 0 we have c(u) T 0 .
R = {v T 0 : h | depthT (v)}.
For any non-leaf vertex u R,
1 2 1 2
w(v)(1 k ) > w(u)(1 k ) .
X
(59)
vDT 0 (u,R)

Recall that DT 0 (, ) is given in Definition 3.3.


For every u, v R such that depthT (v) = depthT (u) + h and v is a descendant of u,
there is one and only one w S such that w lies on the path joining u and v and
depthT (u) < depthT (w) 6 depthT (v).
For any u T 0 such that DT 0 (u, S) 6= , all the vertices of DT 0 (u, S) are at the same
depth in Tu0 , which is an integer between 1 and 2h.
T 0 is sparsified at the subset S.
Lemma 6.5 will be proved in Section 8. Assuming the validity of Lemma 6.5, as well as
the validity of Lemma 6.2 (which will be proved is Section 7), we are now in position to
deduce Lemma 5.5.
Proof of Lemma 5.5. Let F 0 : T 0 2X be the partition map of Lemma 6.2, constructed
with parameters m, h, k, and having the associated designated child map c from Lemma 6.2.
Let T be the tree obtained by applying Lemma 6.5 to (T 0 , w, c), where w : T 0 (0, ) is
given by w(v) = (Fv0 ). Define a fragmentation map F : T 2X by F = F 0 |T , i.e., by
restricting F 0 to the subtree T . Properties {1}, {2} are satisfied by F 0 due to Lemma 6.2,
and therefore they are also satisfied by F since T has the same root as T 0 . Properties {3},
{4}, are part of the conclusion of Lemma 6.5. It remains to prove properties {5} and {6}.

24

Assume that u S. Take y Fr(T ) r (Fu ). In order to prove property {5} it suffices
to show that d(y, Fu ) > 13
2
diam(Fu ). By property {4} it follows that u = c(p(u)). Let
w = lcaT (u, y) and take u0 , y 0 DT (w, S) such that u0 is a weak ancestor of u and y 0 is a
weak ancestor of y. By Lemma 6.5 we know that depthT (u0 ) = depthT (y 0 ), and therefore
by conclusion (56) of Lemma 6.2 and using the fact that c(p(u0 )) = u0 and c(p(y 0 )) = y 0
(because u0 , y 0 S),
1 3 depthT (u0 )1 1 3 depthT (u) (37) 1 3
d (y, Fu ) > d (Fy0 , Fu0 ) > > > diam(Fu ).
2 2 2
It remains to prove property {6}. Take q T and let u T be a weak descendent of q
that has at least two children in T , i.e., v, w p1 (u) T , v 6= w. Our goal is to show that
2 2h
diam (Fq ) 6 depthT (q)depthT (u) d (Fv , Fw ) . (60)
1 3
Since v and w are siblings in T we know by {4} that {v, w} S = . Note that
[ [
Fv = Fy and Fw = Fx ,
yD(v,S) xD(w,S)

and therefore
d (Fv , Fw ) = min d (Fy , Fx ) . (61)
yD(v,S)
xD(w,S)

Note that since {v, w} S = we have D(v, S) D(w, S) D(u, S). By Lemma 6.5 it
follows that all the vertices in D(v, S) D(w, S) are at the same depth in T . Denote this
depth by `. Due to Lemma 6.5 we know that ` 6 depthT (u) + 2h. By conclusion (56) of
Lemma 6.2 we deduce that for all y D(v, S) and x D(w, S) we have
 1 3 ` 1 3 depth (u)+2h
d (Fy , Fx ) > d (Fy , Fx ) = d Fc(p(y)) , Fc(p(x)) > > T . (62)
2 2
Now, the desired inequality (60) is proved as follows.
(61)(62) 1 3 depthT (u)+2h (37) 1 3 depthT (u)depthT (q)+2h
d (Fv , Fw ) > > diam(Fq ). 
2 2
7. The initial fragmentation map: proof of Lemma 6.2
Proof of Lemma 6.2. The construction of the initial fragmentation map will be in a bottom-
up fashion: the tree T will be decomposed as a disjoint union on levels V0 , V1 , . . . , Vmh ,
where Vi are the vertices at depth i. We will construct these levels Vi and the mappings
F : Vi 2X and whk : Vi (0, ) by reverse induction on i, and describe inductively for
each v Vi+1 its parent u Vi , as well as the designated child c(u). At the end of this
construction V0 will consist of a single vertex, the root of T .
Define `mh = |X| and write X = {x1 , . . . , x`mh }. The initial level Vmh consists of the leaves
of T , and it is defined to be Vmh = {vjmh }`j=1
mh
. For all j {1, . . . , `mh } we also set
k1 k1
Fvjmh = {xj } and whk (vjmh ) = w(xj ) k = ({xj }) k .
Assume inductively that for i {1, . . . , mh 1} we have already defined
n o
i+1 i+1 i+1
Vi+1 = v1 , v2 , . . . , v`i+1 ,

25
and the mappings F : Vi+1 2X and whk : Vi+1 (0, ).
Choose j1 {1, . . . , `i+1 } such that
whk vji+1 k i+1
 
1
= max w h vj .
j{1,...,`i+1 }

Define    1 3 
Ai1
= s {1, . . . , `i+1 } : d Fvi+1 , Fvsi+1 6 i .
j1 2
Create a new vertex v1i Vi and define
[
Fv1i = Fvsi+1 .
sAi1

Also, declare the vertices {vsi+1 }sAi1 Vi+1 to be the children of v1i , and in accordance
with (52) define ( k1
w (v i ) k if h | i,
whk v1i = P 1

k i+1
sAi w h (vs ) if h - i.
1

Finally, set
c v1i = vji+1

1
.
Continuing inductively, assume that we have defined v1i , v2i , . . . , vzi Vi , together with
nonempty disjoint sets
Ai1 , . . . , Aiz {1, . . . , `i+1 }.
Sz
If t=1 Ait = {1, . . . , `i+1
Sz} then define `i = z and Vi = {v1i , v2i , . . . , vzi }. Otherwise, choose
i
jz+1 {1, . . . , `i+1 } r t=1 At such that
 
whk vji+1 max Sz i whk vji+1 ,

z+1
= (63)
j{1,...,`i+1 }r t=1 At

and define
( z
)
[   1 3 i
Aiz+1 = s {1, . . . , `i+1 } r Ait : d Fvi+1 , Fvsi+1 6 . (64)
t=1
jz+1 2
i
Create a new vertex vz+1 Vi and define
[
Fvz+1
i = Fvsi+1 . (65)
sAiz+1

Also, declare the vertices {vsi+1 }sAiz+1 Vi+1 to be the children of vz+1
i
and define
(  k1
i
k i
 w vz+1 k
if k | i,
wh vz+1 = P k i+1
sAi wh (vs ) if k - i.
z+1

Finally, set
i
= vji+1

c vz+1 z+1
. (66)
The above recursive procedure must terminate, yielding the level i set Vi . We then proceed
inductively until the set V1 has been defined. We conclude by defining V0 to be a single new
vertex r(T ) (the root) with all the vertices in V1 its children. The designated child of the

26
root, c(r(T )), is chosen to be a vertex u V1 such that whk (u) = maxvV1 whk (v). We also
k1
set Fr(T ) = X and whk (r(T )) = (X) k .
The resulting fragmentation map F : T 2X is by definition a partition map, since
F(Vmh ) = F(L(T )) = X. Also, the construction above guarantees the validity of (55) due
to (63) and (66).
We shall now prove (54) by reverse induction on depthT (u). If depthT (u) = mh then
diam(Fu ) = 0 and there is nothing to prove. Assuming the validity of (54) whenever
i
depthT (u) = i + 1, suppose that depthT (u) = i and moreover that u = vz+1 in the above
construction. By virtue of (64) and (65) we know that
 
i+1 1 3 i
6 3 i+1 + (1 3 ) i = i .

diam (Fu ) = diam Fvz+1 i 6 3 max diam F vs
+ 2
sAiz+1 2
Since (54) is also valid for i = 0 (because diam(X) = 1), this concludes the proof of (54).
It remains to prove (56). Since we are assuming that u 6= v are non-leaf vertices and
depthT (u) = depthT (v), we may write u = vsi and v = vti for some i {1, . . . , mh 1} and
s < t. Then by the above construction c (vsi ) = vji+1
s
, c (vti ) = vji+1
t
and
t1
[ s1
[
jt {1, . . . , `i+1 } r Ai` {1, . . . , `i+1 } r Ai` ,
`=1 `=1

yet jt
/ Ais . The validity of (56) now follows from the definition of Ais ; see (64). 

8. An iterated Ho
lder argument for trees: proof of Lemma 6.5
Our goal here is to prove Lemma 6.5. The heart of this lemma is the extraction of
a large and sparsified subtree from any subadditive weighted tree with designated
children (Definition 6.3). The resulting tree is depicted in Figure 4.

R, S 0

S
R h

S
R 2h
S

R 3h

S
R 4h

Figure 4. A schematic depiction of the subtree T 0 constructed in


Lemma 6.5. The vertices of R are on the dotted lines and the vertices
of S are on the curved solid lines. The vertices of S are -separated.
This is achieved by pruning all their siblings, leaving each of them as
the single offspring of its parent.

27
Definition 8.1 (Sparsified tree). Fix integers h, k > 2 and let (T, w, c) be a subadditive
weighted tree with designated children. For i Z define a subtree T (i) of T as follows.

[ [
T (i) = T r
Tv
. (67)
uT vp1 (u)r{c(u)}
depthT (u)=i1

Thus T (i) is obtained from T by removing all the subtrees rooted at vertices of depth i that
are not designated children. Note that by definition T (i) = T if either i 6 0 or T has no
vertices at depth i.
Lemma 8.2 below is inspired by an argument in [5, Lem. 3.25], though our assumptions,
proof, and conclusion are different.
Lemma 8.2. Fix h, k N satisfying h > k > 2 and let (T, w, c) be a subadditive weighted
tree with designated children. Assume that all the leaves of T are at depth h. Then there
exists L {1, . . . , h} with |L| > h k + 1 such that for every i L we have
k1 k1
X
w(`) k > w(r) k ,
`L(T (i) )

where r = r(T ) is the root of T and T (i) is as in Definition 8.1.


Proof. For i {1, . . . , h} and u T define fi (u) (0, ) by reverse induction on depthT (u)
as follows. If depthT (u) = h, i.e., u is a leaf of T , set
k1
fi (u) = w(u) k . (68)
If depthT (u) < h define recursively

max 1 fi (v) if i = depthT (u) + 1,
fi (u) = P vp (u) (69)
vp1 (u) fi (v) if i 6= depthT (u) + 1.
We observe that for all i {1, . . . , h} and u T we have
k1
X
i 6 depthT (u) = fi (u) = w(`) k , (70)
`L(Tu )

and
k1
X
i > depthT (u) = fi (u) = w(`) k , (71)
`L((Tu )(i) )

where analogously to (67) we define



[ [
(Tu )(i) = Tu r
Tb
. (72)
aT bp1 (a)r{c(a)}
depthT (a)=i1

In other words, recalling that Tu is the subtree of T rooted at u, the subtree (Tu )(i) is obtained
from Tu by deleting all the subtrees rooted at vertices of depth i that are not designated
children (here depth is measured in T , i.e., the distance from the original root r(T )).

28
Identities (70) and (71) follow by reverse induction on depthT (u) from the recursive def-
inition of fi (u). Indeed, if depthT (u) = h then (71) is vacuous and (70) follows from (68).
Assume that u T is not a leaf of T and that (70) and (71) hold true for the children of u.
If i 6 depthT (u) then by (69) and the inductive hypothesis we have
k1 k1
X X X X
fi (u) = fi (v) = w(`) k = w(`) k .
vp1 (u) vp1 (u) `L(Tv ) `L(Tu )

If i = depthT (u) + 1 then since we are assuming that (70) holds for each v p1 (u),

(69) X k1 (53)
fi (u) = max
1
fi (v) = max
1
w(`) k = max
1
whk (v)
vp (u) vp (u) vp (u)
`L(Tv )
(58) (53) X k1 (72) X k1
= whk (c(u)) = w(`) k = w(`) k .
`L(Tc(u) ) `L((Tu )(i) )

Finally, if i > depthT (u) + 1 then we are assuming that (71) holds for each v p1 (u), and
therefore
(69) X (71) X X k1 (72)
X k1
fi (u) = fi (v) = w(`) k = w(`) k .
vp1 (u) vp1 (u) `L((Tv )(i) ) `L((Tu )(i) )

This completes the inductive verification of the identities (70) and (71).
Our next goal is to prove by reverse induction on depthT (u) that for every H {1, . . . , h}
with |H| = k we have,
Y
fi (u) > w(u)k1 . (73)
iH

Indeed, if depthT (u) = h then (73) holds as equality due to (68). Assume inductively that
depthT (u) < h and that (73) holds for all the children of u. We claim that there exists j H
such that

Y   Y X
fi (u) > max
1
fj (v) fi (v) . (74)
vp (u)
iH iHr{j} vp1 (u)

Indeed, if depthT (u) + 1 H then take j = depthT (u) + 1 and note that (74) holds as
equality due to (69). On the other hand, if depthT (u) + 1
/ H then let j be an arbitrary
element of H, and note that due to (69) we have

Y X Y X
fi (u) = fj (v) fi (v)
iH vp1 (u) iHr{j} vp1 (u)

  Y X
> max
1
fj (v) fi (v) ,
vp (u)
iHr{j} vp1 (u)

29
as required. Now,
k1
()
 
Y X Y 1
fi (u) > max
1
f j (v) fi (v) k1
vp (u)
iH vp1 (u) iHr{j}
k1 k1
X Y 1 () X (57)
> fi (v) k1 > w(v) > w(u)k1 ,
vp1 (u) iH vp1 (u)

where in () we used (74) and Holders inequality, and in () we used the inductive hypoth-
esis. This concludes the proof of (73).
We are now in position to complete the proof of Lemma 8.2. Set H1 = {1, . . . , k}. Then
k1
! k1
X k1
Y X k1 (71) Y (73) k1
max w(`) k > w(`) k = fi (r) > w(r) k .

iH1
`L(T (i) ) iH1 `L(T (i) ) iH1

Hence there is i1 H1 satisfying


k1 k1
X
w(`) k > w(r) k .
`L(T (i1 ) )

Now define H2 = (H1 r {i1 }) {k + 1} and repeat the above argument with H2 replacing
H1 . We deduce that there exists i2 H2 satisfying
k1 k1
X
w(`) k > w(r) k .
`L(T (i2 ) )

We may repeat this process inductively h k + 1 times, and let L = {i1 , i2 , . . . , ihk+1 }. 
Proof of Lemma 6.5. The proof is by induction on the height of the tree, but we need the
following strengthening of the inductive hypothesis so as to deal with multiple trees. Suppose
that we are given a collection of (disjoint) subadditive weighted trees with designated children
(T1 , w, c) . . . , (T` , w, c), each Ti is rooted at ri and all of them having the same height, which
is divisible by h (formally we should denote the weighting of Ti by wi , but since the trees
are disjoint, denoting all the weightings by w will not create any confusion). We will prove
that there exists a subset C {1, . . . , `} with the following properties.
For every i C there is a subtree Ti0 of Ti rooted at ri , and subsets Si , Ri Ti0 , both
containing the root of Ti0 , such that for any non-leaf u Ti0 we have c(u) Ti0 .
Ri = {v Ti0 : h | depthTi (v)}.
For any non-leaf vertex u Ri ,
1 2 1 2
w(v)(1 k ) > w(u)(1 k ) .
X
(75)
vDT 0 (u,Ri )
i

For every u, v Ri such that depthTi (v) = depthTi (u) + h and v is a descendant of
u, there is one and only one w Si such that w lies on the path joining u and v and
depthTi (u) < depthTi (w) 6 depthTi (v).

30
For any u Ti0 such that DTi0 (u, Si ) 6= , all the vertices of DTi0 (u, Si ) are at the same
depth in (Ti0 )u , which is an integer between 1 and 2h.
Ti0 is sparsified at
Sthe subset Si .
The vertices in iC DTi0 (ri , Si ) have the same depth (in their respective tree), re-
gardless of i, and
`
!1 k1
1 2
w(r )(1 k ) >
1
X X
i w(r )1 k . i (76)
iC i=1

Note that Lemma 6.5 is the case ` = 1 of this statement, but it will be beneficial to prove
the more general statement as formulated above.
When the height of all the Ti is 0 we simply set C = {1, . . . , `} and Ti0 = Si = Ri = Ti .
Most of the above conditions hold vacuously in this case (there are no non-leaf vertices and
1
DTi0 (u, Si ) = ). Condition (76) follows from subadditivity of the function (0, ) 3 t 7 t1 k .
Assume next that the leaves of {Ti }`i=1 are all at depth mh for some m N. Let Tbj be the
subgraph of Tj induced on all the vertices of depth at most h in Tj . Note that by (53) the
k
restriction of the induced weight function  wh to Tj coincides with the corresponding weight
b
function induced by the weighted tree Tbj , w , and consequently the same can be said about
 
the designated child map c|Tbj . Therefore, an application of Lemma 8.2 to Tbj , w, c yields
a subset Lj {1, . . . , h} with |Lj | = h k + 1 such that for all i Lj we have
1 1
X
w(u)1 k > w(rj )1 k , (77)
(i)
uTbj
depthTj (u)=h

(i)
where Tbj is the subtree of Tbj that is obtained by sparsifying the ith level as in Definition 8.1.
Let j0 {1, . . . , `} satisfy
w(rj0 ) = max w(rj ). (78)
j{1,...,`}

For j {1, . . . , `} denote L0j = Lj Lj0 . Then


|L0j | = |Lj | + |Lj0 | |Lj Lj0 | > (h k + 1) + (h k + 1) h = h 2(k 1). (79)
If ` = 1 let s0 Lj0 be an arbitrary integer in Lj0 . If ` > 2 let s0 Lj0 satisfy
1 1
X X
w(rj )1 k = max w(rj )1 k . (80)
sLj0
j{1,...,`}r{j0 } j{1,...,`}r{j0 }
s0 L0j sL0j

By averaging we see that


X 1 1 X X 1
w(rj )1 k > w(rj )1 k
h k + 1 sL
j{1,...,`}r{j0 } j0 j{1,...,`}r{j0 }
s0 L0j sL0j

1 X 1 (79) h 2(k 1) X 1
= |L0j |w(rj )1 k > w(rj )1 k . (81)
hk+1 h (k 1)
j{1,...,`}r{j0 } j{1,...,`}r{j0 }

31
Now define
C = {j {1, . . . , `} : s0 Lj } . (82)
0
We know that j0 C, since by construction s0 Lj0 . By (79) we have Lj 6= for every
j {1, . . . , `}. Therefore (80) implies that if ` > 2 then |C| > 2. Since (76) is trivial when
` = 1, we will now prove (76) assuming ` > 2. By the choice of j0 in (78) we know that for
all j {1, . . . , `} r {j0 } we have
`
1 1X 1
w (rj )1 k 6 w(ri )1 k . (83)
2 i=1
This implies that for all j {1, . . . , `} r {j0 },
1
1 2 h (k 1) w(rj )1 k
w(rj )(1 k ) >  1 . (84)
h 2(k 1) P` 1 k1 k
i=1 w(ri )

To check (84) note that it is equivalent to the inequality


1
! k1
h (k 1) w(rj )1 k
P` 1 6 1.
h 2(k 1) i=1 w(ri )1 k
By (83) and the fact that h > 2k 2 , it suffices to show that (2k 2 k + 1)/(2k 2 2k + 2) 6 21/k .
Since 21/k > 1 + 1/(2k) it suffices to check that 2k(2k 2 k + 1) 6 (2k + 1)(2k 2 2k + 2),
which is immediate to verify.
Having established (84), we proceed as follows.
1
( 1 k1 )
2 w(rj0 )1 k 1 2
w(rj )(1 k )
X X
w(rj ) = 1 1
+
jC w(rj0 )(1 k ) k jCr{j0 }
1 1 k1
P
w(rj0 )1 k h (k 1) jCr{j0 } w(rj )
(84)
> P  1 + h 2(k 1) P 1
` 1 k1 k ` 1 k1 k
i=1 w(ri ) i=1 w(ri )
1 1 k1
P
w(rj0 )1 k j{1,...,`}r{j0 } w(rj )
(81)(82)
> P  1 + P 1
` 1 k1 k ` 1 k1 k
i=1 w(r i ) i=1 w(r i )
!1 k1
`
1
X
= w(ri )1 k ,
i=1

completing the proof of (76).


We can now complete the proof of Lemma 6.5 by applying the inductive hypothesis. For
(s )
every j C let uj1 , uj2 , . . . , uj`j be the leaves of the tree Tbj 0 , i.e., the subtree of Tbj that
was sparsified at level s0 . Consider the subtrees of Tj that are rooted at uj1 , uj2 , . . . , uj`j , i.e.,
(Tj )uj , (Tj )uj , . . . , (Tj )uj . By the inductive hypothesis applied to these trees there exists a
1 2 `j

subset Cj {1, . . . , `j } such that for each i Cj there is a subtree (Tj )0uj of (Tj )uj and
i i
subsets Sji , Rji (Tj )0uj that satisfy the inductive hypotheses.
i

32
Denote for j C,

[ [ [n (s )
o
Tj0 = (Tj )0uj u Tbj 0 : u ancestor of uji .
i
iCj iCj

(s )
Thus Tj0 is obtained by taking the subtree of Tbj 0 whose leaves are uji iCj , and replacing

S S
j 0 `j
every leaf ui by the tree (Tj )uj . We also define Rj = i=1 Rji {rj }, and
i

`j
[   [ n (s )
o[
Sj = Sji r uji u Tbj 0 : depthTj (u) = s0 {rj }.
i=1

(s ) (s ) (s )
Note by the definition of Tbj 0 every u Tbj 0 with depthTj (u) = s0 has no siblings in Tbj 0 .
All the desired properties of Tj0 , Rj , Sj follow immediately for the construction; only (75)
when u = rj requires justification as follows.
1 k1
2 `j
1 2 1
(1 k ) (76) 1
1 k (77) 1 2
w(v)(1 k ) = > w(rj )(1 k ) . 
X X X
w uji > w uji
vDT 0 (rj ,Rj ) iCj i=1
j

9. Proof of Theorem 5.3


Here we prove Theorem 5.3, which is the last missing ingredient of the proof of Theo-
rem 1.5. Theorem 5.3 is a weighted version of the results of [25, 30]. Theorem 9.1 below
follows from a slight modification of the argument in [30], though it is not stated there
explicitly. We will therefore explain how [30] can be modified to deduce this statement.
Alternatively, a similar statement (with worse distortion bound) can be obtained by natural
modifications of the argument in [25].
Theorem 9.1. Let (X, d) be a finite metric space and w1 , w2 : X [0, ) two nonnegative
weight functions. Then for every (0, 1) there exists a subset S X that embeds into an
ultrametric space with distortion
2
D= 1 , (85)
(1 )
and satisfying ! !
X X X
w1 (x) w2 (x) > w1 (x)w2 (x) . (86)
xS xX xX

Theorem 9.1 implies the finite nonlinear Dvoretzky theorem (Theorem 1.2) in the special
case w1 = w2 = 1. Theorem 5.3 follows by taking w1 = w1 and w2 = w. In this case
conclusion (86) becomes
!1
X X
w(x)1 > w(x) . (87)
xS xX
This type of requirement was studied in [5] under the name of the weighted metric Ramsey
problem, where is was shown that there always exists S X satisfying (87) that embeds
into an ultrametric space with distortion O (1 log(2/)).

33
Proof of Theorem 9.1. The beginning of the argument is most natural to state in the context
of general metric measure spaces (X, d, ). So, assume that (X, d, ) is a metric measure
space; we will later specialize the discussion to the case of finite spaces.
Let f : X [0, ) be a nonnegative Borel measurable function. Lemma 2.1 of [30] states
that for every compact S X and every R > r > 0 there exists a compact subset T S
satisfying
(B(x, R))
Z Z
f (x)d(x) > f d, (88)
T (B(x, r)) S
such that T can be partitioned as T =
S
n=1 Tn , where each (possibly empty) Tn is compact
and contained in a ball of radius r, and any two non-empty Tn , Tm are separated by a distance
of at least R r.
Fix a nonnegative Borel measurable w L1 (). Iterate the above statement as follows;
the same iteration is carried out for the special case w = 1 in Lemma 2.2 of [30]. Assume
that we are given a non-increasing sequence of positive numbers R = r0 > r1 > r2 > > 0
converging to zero. Assume also that diam(X) 6 2R. For n N define fn : X [0, ) by

!
Y (B(x, rm ))
fn (x) = 2rm1
 w(x), (89)
m=n
B x, rm + D

Where D be given by (85). Note that 0 6 fn 6 w for all n N. Assume that we already
defined a compact subset Sn1 X. An application of Lemma 2.1 of [30], with radii
rn + 2rD
n1
> rn and weight function fn , yields a compact subset Sn Sn1 satisfying
B x, rn + 2rD
Z Z n1
 Z
(88)
fn+1 d = fn (x)d(x) > fn d.
Sn Sn (B(x, rn )) Sn1

Hence for all n N we have, Z Z


wd > f1 d.
TSn X
Consider the compact subset S = n=1 Sn . By the dominated convergence theorem,
Z Z
wd > f1 d. (90)
S X
In [30, Lem. 2.2] it is shown that S embeds with distortion D into an ultrametric space.
Assume now that the radii 1 = r0 > r1 > r2 > > 0 are random variables satisfying
limn rn = 0 and for every real number r > 0
 
2rn1
Pr rn < r 6 rn + 6 . (91)
D
For the existence of such random variables, as well as the optimality for this purpose of the
choice of D in (85), see [30, Thm. 1.5]. Specializing to the case of a finite metric measure space
(X, d, ) of diameter at most 2, apply (90) when w(x) = w1 (x)/({x}) and w2 (x) = ({x}),
and the radii are the random radii chosen above. By taking expectation of the resulting
(random) inequality and using Jensens inequality we arrive at the following estimate.
" # " !#!
X (89)(90) X X (B(x, rm ))
E w1 (x) > w1 (x) exp E log . (92)
B x, rm + 2rm1

xS xX n=1 D

34
For every x X let 0 = t1 (x) < t2 (x) < . . . < tk(x) (x) be the radii at which (B(x, t))
jumps, i.e., ({x}) = (B(x, t1 (x))) < (B(x, t2 (x))) < . . . < (B(x, tk(x) (x))) = (X), and
B(x, t) = B(x, tj (x)) if tj (x) 6 t < tj+1 (x) (where we use the convention tk(x)+1 (x) = ).
Then we have the following straightforward identity (see equation (15) in [30]), which holds
for every x X.
" !#
X (B(x, rm ))
E log 2rm1

n=1
B x, rm + D
k(x) !
X X  2rn1

(B(x, tj (x)))

= Pr rn < tj (x) 6 rn + log . (93)
j=2 n=1
D (B(x, tj1 (x)))
Hence,

k(x)
" #  
X (91)(92)(93) X X (B(x, tj (x)))
E w1 (x) > w1 (x) exp log
xS xX j=2
(B(x, tj1 (x)))
 P
w1 (x)w2 (x)

X ({x})
= w1 (x) = xX  .
xX
(X) P
w (y)
yX 2

We have shown that the required estimate (86) holds in expectation for our random subset
S X, completing the proof of Theorem 9.1. 

10. Impossibility results


The purpose of this section is to prove the second part of Theorem 1.4 and Theorem 1.8.
In both cases the goal is to construct a metric space having the property that all its almost
Euclidean subsets have small Hausdorff dimension. We will do so by gluing together the
finite examples from [5]: in the high distortion regime corresponding to Theorem 1.4 these
building blocks are expander graphs, and in the low distortion regime corresponding to The-
orem 1.8 these building blocks arise from dense random graphs. The gluing procedure, which
is an infinitary variant of the metric composition method from [5], starts with a sequence
of finite metric spaces and joins them in a tree-like fashion. The details of the construc-
tion are contained in Section 10.1 below, and the specializations to prove Theorem 1.4 and
Theorem 1.8 are described in Section 10.2 and Section 10.3, respectively.
10.1. Trees of metric spaces. Fix {nk } k=0 N with n0 = 1 and nk > 1 for k > 1. Fix
also {k }
k=1 (0, ). Assume that for each k N we are given a metric dk on {1, . . . , nk }
with
diamdk ({1, . . . , nk }) = 1 and min dk (i, j) = k . (94)
i,j{1,...,nk }
Q
For distinct x = (xk )
k=1 , y = (yk )k=1 k=1 {1, . . . , nk } let k(x, y) be the smallest k N
such that xk 6= yk . For (0, ) define

dk(x,y) xk(x,y) , yk(x,y)
(x, y) = Qk(x,y)1 1/ . (95)
i=0 n i

Also, set (x, x) = (y, y) = 0.

35
Remark 10.1. One can visualize the above construction as follows. Let T be the infinite
rooted
Q tree such for i > 0 each vertex at depth i in T has exactly ni+1 children. Then
k=1 {1, . . . , nk } can be identified with the set of all infinite branches of T . With this
identification, the distance has the following meaning: given two infinite branches in T ,
find the vertex v T at which they split (i.e., their deepest common vertex). Say that the
depth of v is i 1. The metric di induces a metric space structure on the ni children of v,
and the distance between the two given branches is a multiple of the distance between the
two children of v that belong to these branches.
Lemma 10.2. (
Q
k=1 {1, . . . , nk }, ) is a compact metric space provided that
1
k N, k > 1/
. (96)
nk
Proof. Take x, y, z
Q
k=1 {1, . . . , nk }. If k(x, y) = k(y, z) = k(x, z) = k then because
dk satisfies the triangle inequality, (x, z) 6 (x, y) + (y, z). If k(x, y) > k(x, z) then
necessarily k(x, z) = k(y, z) = k and xk = yk . Hence (x, z) = (x, y) 6 (x, y)+ (y, z).
The remaining case k(x, z) > k(x, y) is dealt with as follows.

dk(x,z) xk(x,z) , zk(x,z) (94) 1 (96) k(x,y) (94)
(x, z) = Qk(x,z)1 1/ 6 Qk(x,z)1 1/ 6 Qk(x,y)1 1/ 6 (x, y).
i=0 ni i=0 ni i=0 ni

Q Compactness follows from Tychonoffs theorem since


This proves the triangle inequality.
induces the product topology on k=1 {1, . . . , nk }. 
Lemma 10.3. Assume that in addition to (96) we have
log(1/k )
lim Pk1 = 0. (97)
i=1 log ni
k

Then

!
Y
dimH {1, . . . , nk }, = .
k=1

Proof. Define = dimH (


Q
k=1 {1, . . . , nk }, ). The fact that 6 is simple. Indeed, for
k
k N consider the sets {Bx }xQk {1,...,ni } given by
i=1

k
!
!
Y Y
k
Bx = {xi } {1, . . . , ni } . (98)
i=1 i=k+1
1/
and {Bxk }xQk {1,...,ni } cover
Qk
Then diam (Bxk ) = i=1 ni
Q
Q i=1 i=1 {1, . . . , ni }. Hence for
> the -Hausdorff content of ( i=1 {1, . . . , ni }, ) can be estimated as follows.

!

Y X 1 1
H {1, . . . , ni }, 6 inf Qk /
= inf Qk ()/
= 0.
i=1
kN
x
Qk
{1,...,n } i=0 n i
kN
i=0 n i
i=1 i

Thus 6 .
We now pass to the proof of > . We first prove the following preliminary state-
Q
1 m
ment. Assume that x , . . . , x i=1 {1, . . . , ni } and k1 , . . . , km N {0} are such

36
k
that {Bxjj }m cover k
Q
j=1 i=1 {1, . . . , ni }, where Bx is given in (98) and we use the convention
Bx0 =
Q
i=1 {1, . . . , ni }. We claim that this implies that
m
X 1
Qk j > 1. (99)
j=1 i=0 ni
The proof is by induction on m. If m = 1 then k1 = 0 and (99) follows. Assume that
k Q
m > 2, no subset of {Bxjj }m j=1 covers i=1 {1, . . . , ni }, and that k1 6 k2 6 6 km .
For every y {1, . . . , nkm } let x (y)
km
Q
i=1 {1, . . . , ni } have y in the km th coordinate,
and coincide with xkm in all other coordinates. The sets {Bxkkmm (y) }y{1,...,nkm } are pair-
k
wise disjoint, and are either contained in or disjoint from Bxjj for each j {1, . . . , m}.
k
Hence, by the minimality of the cover {Bxjj }m j=1 we have km = km1 = = kmnkm +1 and
kj km
{Bxj }j=mnkm +1 = {Bxkkmm (y) }y{1,...,nkm } . This implies that
m mnkm mnkm
X 1 X 1 X 1 X 1 1
Qkj = Qkj + Qkm = Qkj + Qkm 1 .
j=1 i=0 ni j=1 i=0 ni y{1,...,nkm } i=0 ni j=1 i=0 ni i=0 ni
k
m 1 mn
The induction hypothesis applied to {Bxjj }j=1 km {Bxkm } concludes the proof of (99).
Fix (0, ). Due to (97) there exists C (0, ) such that for all k N,
k1
1 Y ()/2
6C ni . (100)
k i=1

Let {B (x , rj )}jJ be a family of balls that covers


j
Q
i=1 {1, . . . , ni }. We will show that
X 1
rj > . (101)
jJ
C

Q
This would mean that QH ( i=1 {1, . . . , ni }, ) > 0 for all (0, ), proving that > .
By compactness of ( i=1 {1, . . . , ni }, ) it suffices to prove (101) when J is finite. For every
Qkj 1/ Qkj 1 1/
j J choose kj N such that i=0 ni < rj 6 i=0 ni . Define
( Q
kj 1 1/ Q kj 1 1/ Qkj 1 1/
i=1 ni if kj i=1 ni 6 rj 6 i=1 ni ,
rj = Qkj 1/ Qkj 1/ Qkj 1 1/ (102)
n
i=1 i if n
i=1 i < r j < k j i=1 ni .
Qkj 1 1/ Qkj 1 1/ k 1
If kj i=1 ni 6 rj 6 i=1 ni then rj > rj , hence B (xj , rj ) B (xj , rj ) = Bxjj .
Q Qkj 1 1/ 
kj 1/
Also, observe that does not take values in the interval i=1 ni , kj i=1 ni .
k k 1/ k 1 1/
This implies that B (xj , rj ) = B (xj , rj ) = Bxjj when i=1
Q j
Q j
ni < rj < kj i=1 ni .
j
Q
We deduce that the balls B (x , rj ) cover i=1 {1, . . . , ni }, and they are all sets of the form
Bxk . It therefore follows from (99) that
j 1
X  (102) X  rj  (100) X kY ()/
X

16 rj 6 6 C rj ni 6 C rj , (103)
jJ jJ
kj jJ i=0 jJ
Qkj 1 1/
where in the last inequality of (103) we used the fact that rj 6 i=0 ni . 

37
Remark 10.4.
Qk A less direct way to prove the bound > in Lemma 10.3 is to define
k 1
(Bx ) = i=0 ni and to argue that the Caratheodory Q extension theorem applies here
and yields an extension of to a Borel measure on i=1 {1, . . . , ni }.Q One can then show
analogously to (103) that this measure is a -Frostman measure for ( i=1 {1, . . . , ni }, ).

In what follows we say that a property P of metric spaces is a metric property if whenever
(X, dX ) P and (Y, dY ) is isometric to (X, dX ) then also (Y, dY ) P. We say that P is
hereditary if whenever (X, d) P and Y X then also (Y, d) P. Finally, we say that
P is dilation-invariant if whenever (X, d) P and (0, ) also (X, d) P.
Theorem 10.5. Fix > 0. Fix also {nk } k=0 N with n0 = 1 and nk > 1 for k > 1, a
sequence {k }k=1 (0, ), and for each k N a metric dk on {1, . . . , nk }. Assume that (94),
(96) and (97) hold true. Then there exists a metric space (Y, ) with dimH (Y, ) = that
satisfies the following property. Let {Pk } i=1 is a non-decreasing (with respect to inclusion)
sequence of hereditary dilation-invariant metric properties and for every k N let mk be the
cardinality of the largest subset S ofS{1, . . . , nk } such that (S, dk ) has the property Pk . Then
every Z Y that has the property k=1 Pk satisfies

log mk
dimH (Z, ) 6 lim sup .
k log nk
Q
Proof. Take (Y, ) = ( k=1 {1, . . . , nk }, ), where is given in (95). By Lemma 10.3 and
Lemma 10.2 we know that (Y, ) is a compact metric space of Hausdorff dimension .
Assume that Z Y and (Z, ) PK for some K N. Since {Pk } i=1 are non-decreasing
properties, we know that (Z, ) Pk for all k > K. For every x Y and k N denote
( k1
!
!! )
Y Y
Sxk = j {1, . . . , nk } : Z {xi } {j} {1, . . . , ni } 6= .
i=1 i=k+1

If j Sxk choose xk (j) Z whose first k 1 coordinates coincide with the corresponding
coordinates of x, and whose kth coordinate equals j. Then {xk (j)}jSxk is a subset of Z whose
metric is isometric to a dilation of the metric dk on Sxk . Since Pk is a hereditary dilation-
invariant metric property, it follows that Sxk Pk for all k > K. Hence |Sxk | 6 mk . Let Zk be
the projection of the set Z onto the first k coordinates, i.e., the set of all x ki=1 {1, . . . , ni }
Q
such that Bxk Z 6= , where Bxk is given in (98). Then it follows by induction that for every
k > K we have |Zk | 6 K1
Q Qk log mk
i=0 ni i=K mi . Denote = lim supk log nk . If > then
(+)/(2)
there exists K 0 > K such that for every k > K 0 we have mk 6 nk . Since the sets
1/
{Bxk }xZk cover Z and have -diameter ki=0 ni
Q
,
QK1 Qk

X 1 i=0ni i=K mi
H (Z, ) 6 inf 0 Qk /
6 inf 0 Qk /
i=0 ni i=0 ni
k>K k>K
xZk
QK 0 1 (+)/(2) QK 0 1
ni ki=K 0 ni
Q
i=0 i=0 ni
6 inf 0 Qk /
6 inf k ()/(2)
= 0.
k>K 0
Q
k>K n
i=0 i i=K 0 n i

Hence dimH (Z, ) 6 . 

38
Corollary 10.6. Fix an integer n > 2. Let (X, d) be an n-point metric space, and assume
that (0, ) satisfies
min d(x, y) > diam(X).
x,yX
x6=y
Then there exists a compact metric space (Y, ) with dimH (Y, ) = log n that has the follow-
ing property. Fix m {1, . . . , n 1} and assume that P is a hereditary dilation-invariant
metric property such that the largest subset of X having the property P is of size m. Then
dimH (Z, ) 6 log m for every Z Y with property P.
Proof. By rescaling assume that diam(X) = 1. Now apply Theorem 10.5 with Xi = X,
ni = n, Pi = P and = log n. 
10.2. Expander fractals. It is shown in [5] that there is c (0, ) such that for any
n N there exists an n-point metric space Xn such that for every (0, 1) all the subsets
of Xn of cardinality greater than n1 incur distortion greater than c/ in any embedding
into Hilbert space. In fact, the spaces Xn are the shortest-path metrics on expander graphs,
implying that diam Xn 6 C log n for some C (0, ) and all n N (see [12]). We will
apply Corollary 10.6 to these spaces, thus obtaining compact metric spaces that can be called
expander fractals. The property P that will be used is X embeds with distortion c/
into Hilbert space, which is clearly a hereditary dilation-invariant metric property.
Proof of the second part of Theorem 1.4. Let c, C, Xn be as above. Fix > 0 and choose an
integer n > 2 such that n1/ > C log n. We may therefore use Corollary 10.6 with X = Xn
and = n1/ . The resulting compact metric space (Y, ) will then have Hausdorff dimension
equal to . For every (0, 1) let P be the property X embeds with distortion c/ into
Hilbert space. Then all the subsets Z of Y that embed into Hilbert space with distortion
c/ satisfy dimH (Z, ) 6 log (n1 ) = (1 ) = (1 ) dimH (Y, ). 
10.3. G (n, 1/2) fractals. It is shown in [5] that there exists K (1, ) such that for any
n N there exists an n-point metric space Wn such that for every (0, 1) any subset
2
of Wn of size larger than 2 log2 n + K ( 2 log(2/)) must incur distortion at least 2
when embedded into Hilbert space. The space Wn comes from a random construction:
consider a random graph G on n vertices, drawn from the Erdos-Reyni model G(n, 1/2)
(thus every edge is present independently with probability 1/2). The space Wn is obtained
from G by declaring two vertices that are joined by an edge to be at distance 1, and two
distinct vertices that are not joined by an edge are declared to be at distance 2. Therefore
the positive distances in Wn are either 1 or 2. This description of Wn is implicit in [5]
but follows immediately from the proof of [5]; see [6] for an alternative proof of this fact
(yielding a worse asymptotic dependence on that is immaterial for our purposes). We will
apply Theorem 10.5 to {Wn } n=2 , thus obtaining compact metric spaces that can be called
G(n, 1/2) fractals.
Proof of Theorem 1.8. Let K and {Wn } n=2 be as in the above discussion. Set Xn = Wn+d3 e .
Hence |Xn | > 2. Let Pn be the property X embeds with distortion 2 1/ log log n into
1/

Hilbert space. Then {Pn }n>20Sis a non-decreasing sequence of hereditary dilation-invariant


metric properties. Moreover, n>20 Pn is the property X embeds into Hilbert space with
distortion smaller than 2. By the above discussion, letting mn be the size of the largest
subset of Xn that has the property Pn , we have mn 6 2 log2 n + O ((log log)2 (log log log n)2 ).

39
Hence lim supn loglog|Xmnn| = 0. By Theorem 10.5 it follows there exists a compact metric
space (Y, ) with Hausdorff dimensionS such that all of its subsets with positive Hausdorff
dimension do not have property n>20 Pn , namely any embeding of such a subset into
Hilbert space must incur distortion at least 2. 

Acknowledgements. We are grateful to Terence Tao for sharing with us his initial attempts
to solve Question 1.3. We thank Tamas Keleti, Andras Mathe and Ondrej Zindulka for help-
ful comments. We are also grateful to an anonymous referee who suggested a reorganization
of our proof so as to improve the exposition.

References
[1] G. Alberti, M. Csornyei, and D. Preiss. Structure of null sets in the plane and applications. In European
Congress of Mathematics, pages 322. Eur. Math. Soc., Z urich, 2005.
[2] L. Ambrosio and B. Kirchheim. Rectifiable sets in metric and Banach spaces. Math. Ann., 318(3):527
555, 2000.
[3] Y. Bartal, B. Bollob as, and M. Mendel. Ramsey-type theorems for metric spaces with applications to
online problems. J. Comput. System Sci., 72(5):890921, 2006.
[4] Y. Bartal, B. Bollob as, and M. Mendel. Ramsey-type theorems for metric spaces with applications to
online problems. J. Comput. System Sci., 72(5):890921, 2006.
[5] Y. Bartal, N. Linial, M. Mendel, and A. Naor. On metric Ramsey-type phenomena. Ann. of Math. (2),
162(2):643709, 2005.
[6] Y. Bartal, N. Linial, M. Mendel, and A. Naor. Some low distortion metric Ramsey problems. Discrete
Comput. Geom., 33(1):2741, 2005.
[7] Y. Benyamini and J. Lindenstrauss. Geometric nonlinear functional analysis. Vol. 1, volume 48 of
American Mathematical Society Colloquium Publications. American Mathematical Society, Providence,
RI, 2000.
[8] A. Blum, H. Karloff, Y. Rabani, and M. Saks. A decomposition theorem for task systems and bounds
for randomized server problems. SIAM J. Comput., 30(5):16241661 (electronic), 2000.
[9] J. Bourgain, T. Figiel, and V. Milman. On Hilbertian subsets of finite metric spaces. Israel J. Math.,
55(2):147152, 1986.
[10] D. Burago, Y. Burago, and S. Ivanov. A course in metric geometry, volume 33 of Graduate Studies in
Mathematics. American Mathematical Society, Providence, RI, 2001.
[11] L. Carleson. Selected problems on exceptional sets. Van Nostrand Mathematical Studies, No. 13. D. Van
Nostrand Co., Inc., Princeton, N.J.-Toronto, Ont.-London, 1967.
[12] F. R. K. Chung. Diameters and eigenvalues. J. Amer. Math. Soc., 2(2):187196, 1989.
[13] A. Dvoretzky. Some results on convex bodies and Banach spaces. In Proc. Internat. Sympos. Linear
Spaces (Jerusalem, 1960), pages 123160. Jerusalem Academic Press, Jerusalem, 1961.
[14] K. Funano. Two infinite versions of nonlinear Dvoretzkys theorem, 2011. Preprint available at http:
//arxiv.org/abs/1111.1627.
[15] A. Grothendieck. Sur certaines classes de suites dans les espaces de Banach et le theor`eme de Dvoretzky-
Rogers. Bol. Soc. Mat. S ao Paulo, 8:81110 (1956), 1953.
[16] J. D. Howroyd. On dimension and on the existence of sets of finite positive Hausdorff measure. Proc.
London Math. Soc. (3), 70(3):581604, 1995.
[17] B. Hughes. Trees and ultrametric spaces: a categorical equivalence. Adv. Math., 189(1):148191, 2004.
[18] H. Karloff, Y. Rabani, and Y. Ravid. Lower bounds for randomized k-server and motion-planning
algorithms. SIAM J. Comput., 23(2):293312, 1994.
[19] T. Keleti. A peculiar set in the plane constructed by Vituskin, Ivanov and Melnikov. Real Anal. Ex-
change, 20(1):291312, 1994/95.
[20] T. Keleti, A. Mathe, and O. Zindulka. Hausdorff dimension of metric spaces and Lipschitz maps onto
cubes. Preprint, available at http://arxiv.org/abs/1203.0686, 2012.

40
[21] M. Laczkovich. Paradoxical decompositions using Lipschitz functions. Real Anal. Exchange, 17(1):439
444, 1991/92.
[22] J. R. Lee and A. Naor. Extending Lipschitz functions via random metric partitions. Invent. Math.,
160(1):5995, 2005.
[23] J. Matousek. On Lipschitz mappings onto a square. In The mathematics of Paul Erd os, II, volume 14
of Algorithms Combin., pages 303309. Springer, Berlin, 1997.
[24] P. Mattila. Geometry of sets and measures in Euclidean spaces, volume 44 of Cambridge Studies in
Advanced Mathematics. Cambridge University Press, Cambridge, 1995. Fractals and rectifiability.
[25] M. Mendel and A. Naor. Ramsey partitions and proximity data structures. J. Eur. Math. Soc., 9(2):253
275, 2007.
[26] M. Mendel and A. Naor. Ultrametric skeletons. Preprint, available at http://arxiv.org/abs/1112.
3416, 2011.
[27] V. Milman and G. Schechtman. An isomorphic version of Dvoretzkys theorem. II. In Convex geometric
analysis (Berkeley, CA, 1996), volume 34 of Math. Sci. Res. Inst. Publ., pages 159164. Cambridge Univ.
Press, Cambridge, 1999.
[28] V. D. Milman. A new proof of A. Dvoretzkys theorem on cross-sections of convex bodies. Funkcional.
Anal. i Prilozen., 5(4):2837, 1971.
[29] P. Morters and Y. Peres. Brownian motion. Cambridge Series in Statistical and Probabilistic Mathemat-
ics. Cambridge University Press, Cambridge, 2010. With an appendix by Oded Schramm and Wendelin
Werner.
[30] A. Naor and T. Tao. Scale-oblivious metric fragmentation and the nonlinear Dvoretzky theorem, 2010.
Preprint available at http://arxiv.org/abs/1003.4013. To appear in Israel J. Math.
[31] H. Sagan. Space-filling curves. Universitext. Springer-Verlag, New York, 1994.
[32] G. Schechtman. Two observations regarding embedding subsets of Euclidean spaces in normed spaces.
Adv. Math., 200(1):125135, 2006.
[33] C. Sommer, E. Verbin, and W. Yu. Distance oracles for sparse graphs. In 2009 50th Annual IEEE
Symposium on Foundations of Computer Science (FOCS 2009), pages 703712. IEEE Computer Soc.,
Los Alamitos, CA, 2009.
[34] M. Talagrand. Regularity of Gaussian processes. Acta Math., 159(1-2):99149, 1987.
[35] M. Talagrand. The generic chaining. Springer Monographs in Mathematics. Springer-Verlag, Berlin,
2005. Upper and lower bounds of stochastic processes.
[36] M. Talagrand. Upper and Lower Bounds for Stochastic Processes. 2011. Modern Methods and Classical
Problems. Forthcoming book.
[37] M. Thorup and U. Zwick. Approximate distance oracles. J. ACM, 52(1):124 (electronic), 2005.
[38] M. Urba nski. Transfinite Hausdorff dimension. Topology Appl., 156(17):27622771, 2009.
[39] I. A. Vestfrid and A. F. Timan. A universality property of Hilbert spaces. Dokl. Akad. Nauk SSSR,
246(3):528530, 1979.
[40] A. G. Vituskin, L. D. Ivanov, and M. S. Mel0 nikov. Incommensurability of the minimal linear measure
with the length of a set. Dokl. Akad. Nauk SSSR, 151:12561259, 1963.
[41] C. Wulff-Nilsen. Approximate distance oracles with improved query time. Preprint, available at http:
//arxiv.org/abs/1202.2336, 2011.

Mathematics and Computer Science Department, Open University of Israel, 1 University


Road, P.O. Box 808 Raanana 43107, Israel
E-mail address: mendelma@gmail.com

Courant Institute, New York University, 251 Mercer Street, New York NY 10012, USA
E-mail address: naor@cims.nyu.edu

41